Compare commits

...

41 Commits

Author SHA1 Message Date
Martino Russi fc7a0bc2fd feat(unitree_g1): drive SONIC whole-body from a 34-D OpenHLM/pi0.5 VLA
Add a dense 34-D whole-body command path so lerobot-rollout can drive the
G1 directly with an OpenHLM / pi0.5 policy through the SONIC encoder/decoder:

- SonicWholeBodyController: wb.{i}.pos action interface, mode-0 reference with
  a rolling 50-frame trajectory (finite-diff velocities) and first-tick anchor
  init; correct MuJoCo->IsaacLab joint reordering.
- unitree_g1: expose 34-D wb_state.{i}.pos proprio; empty/replay camera feeds
  for image-conditioned policies; Dex3 hand publishing from the grip scalars.
- g1_utils: obs_to_wb34_state + WB action constants.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-20 20:01:48 +02:00
Martino Russi 5f6513551c Merge branch 'main' into feat/unitree_g1_sonic_rebased 2026-07-18 13:23:23 +02:00
Martino Russi 70e157e00f fix ruff 2026-07-18 13:22:53 +02:00
Martino Russi 1837be51bf add 3 point calibration + waist coupling, remote controller and smoothed motion 2026-07-17 17:56:30 +02:00
Martino Russi bedd56eed9 Remove g1_sonic_slider, examples/onnx, and SONIC debugging docs 2026-07-16 14:40:32 +02:00
Martino Russi c165e4df68 Merge branch 'main' into feat/unitree_g1_sonic_rebased 2026-07-16 14:33:10 +02:00
Martino Russi 5e24da483a (add) sonic 3-point teleop, safe startup/shutdown, tested on real g1 2026-07-16 13:38:49 +02:00
Martino Russi 9c54665a76 test 3-point teleop
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-15 18:20:26 +02:00
Martino Russi f6a845c30c Merge branch 'main' into feat/unitree_g1_sonic_rebased 2026-07-15 17:13:50 +02:00
Martino Russi 45e8336854 replace quat operations with scipy 2026-07-15 17:07:09 +02:00
Martino Russi 5046e2df32 fix ruff 2026-07-15 16:42:46 +02:00
Martino Russi 1c88e26c6d clean up sonic-side 2026-07-15 16:40:56 +02:00
Martino Russi 69a3edfa33 fix lint 2026-07-15 16:00:42 +02:00
Martino Russi 2492ce2c29 switch to logging 2026-07-15 15:30:54 +02:00
Martino Russi c8e75da55f Merge remote-tracking branch 'origin/main' into feat/unitree_g1_sonic_rebased 2026-07-15 14:59:53 +02:00
Martino Russi 2eae31ea2b fix(unitree_g1): disable SMPL root-motion anchor to prevent sim instability
Feeding the per-frame SMPL root quaternion into the mode-2 anchor produced
root-acceleration spikes (NaN QACC at DOF 0) mid-episode during replay. Keep the
anchor self-driven until the reference root trajectory is smoothed/rate-matched
(30 Hz dataset -> 50 Hz control).

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-15 14:59:22 +02:00
Martino Russi c997abe739 (fix) keep num of ORTthreads under core count 2026-07-14 18:11:03 +02:00
Martino Russi c73579055e refactor(unitree_g1): drop duplicate keyboard code, clarify smpl sentinel
- Remove unused RawKeyboard/drain_keyboard/process_keyboard from sonic_pipeline
  (dead code duplicating lerobot.utils.keyboard_input); the G1 integration uses
  the joystick path. Drop now-unused sys/select/termios/tty imports.
- Add a comment explaining the smpl.0 presence check is a sentinel for a full
  SMPL window (review question).

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-14 15:09:03 +02:00
Martino Russi 4be438161b style: apply ruff format to sonic_pipeline and smpl_fk
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-14 14:46:05 +02:00
Martino Russi 806d28a883 docs(unitree_g1): add docstrings and comments to sonic_pipeline
Address review feedback that sonic_pipeline.py was dense and hard to read.
Adds a module-level architecture overview plus class and key-function
docstrings (planner subprocess, encoder/decoder, movement state, input
helpers). No behavior change.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-14 14:40:39 +02:00
Martino Russi 573b65ff6b (fix) hardcode smpl_skeleton, remove .npz 2026-07-14 13:07:30 +02:00
Martino Russi bc55713e7c fix relative imports 2026-07-13 18:50:00 +02:00
Martino Russi 4f53c42583 Apply ruff-format
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-10 16:03:28 +02:00
Martino Russi bfced3d149 Silence ruff N817 on scipy Rotation import 2026-07-10 15:59:55 +02:00
Martino Russi 4969813d4e Silence ruff N817 on scipy Rotation import 2026-07-10 15:49:09 +02:00
Martino Russi 1c87ca31a3 remove examples inlcuding npz motion files 2026-07-10 15:47:21 +02:00
Martino Russi 4bcde762cc add heading to SMPL, stream dataset 2026-07-10 15:44:48 +02:00
Martino Russi 943ae78cfe feat(unitree_g1): standalone PICO SMPL publisher + dedup/replay fixes
Add a self-contained rt/smpl publisher in the pico_headset teleoperator
(pico_publisher.py + numpy SMPL FK in smpl_fk.py + vendored skeleton table)
so headset whole-body teleop no longer depends on gear_sonic/torch; only
xrobotoolkit_sdk is needed at the headset.

Also: share lowstate_to_obs/get_gravity_orientation via g1_utils (dedup
sonic_pipeline and UnitreeG1.get_observation), and fix dataset-replay joint
ordering (Unitree -> IsaacLab) for sonic.py --replay-dataset.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-09 19:13:22 +02:00
Martino Russi 3363688f1e Merge branch 'main' into feat/unitree_g1_sonic_rebased 2026-07-09 18:02:53 +02:00
Martino Russi 0876629e72 Merge branch 'main' into feat/unitree_g1_sonic_rebased 2026-07-06 18:21:16 +02:00
Martino Russi 305614b8c6 add pico teleoperator, add sonic VR support
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-06 18:16:12 +02:00
Martino Russi 02d3202c4f add SMPL wiring into sonic controller 2026-07-06 18:13:46 +02:00
Martino Russi 3b6de2fdf8 fix(unitree_g1): fix typo flagged by spellchecker in motion_loader docstring 2026-06-26 13:46:33 +02:00
Martino Russi 744f3667c0 fix(unitree_g1): silence bandit findings in SONIC example/pipeline 2026-06-26 13:40:53 +02:00
Martino Russi fdde436776 Merge branch 'main' into feat/unitree_g1_sonic_rebased 2026-06-26 13:35:28 +02:00
Martino Russi 5c683c65c6 Merge branch 'main' into feat/unitree_g1_sonic_rebased 2026-06-25 14:38:48 +02:00
Martino Russi dfbc25c58f fix(unitree_g1): satisfy ruff lint/format and address review comments 2026-06-25 14:37:44 +02:00
Martino Russi 804c76bcc2 Merge branch 'main' into feat/unitree_g1_sonic_rebased 2026-06-25 13:41:04 +02:00
Martino Russi e6afa69be9 add motion loader 2026-06-17 12:31:08 +02:00
Martino Russi 31d1439e29 add custom motion loader 2026-06-17 12:29:36 +02:00
Martino Russi 1c118c6359 feat(unitree_g1): add SONIC whole-body controller
Move GrootLocomotionController and HolosomaLocomotionController into a new
controllers/ subpackage and add the SONIC whole-body controller
(sonic_pipeline.py, sonic_whole_body.py) plus the examples/unitree_g1/sonic.py
standalone script. UnitreeG1 now honors a controller's kp/kd, calls
controller.shutdown() on disconnect, and skips arm publishing for full_body
controllers.
2026-06-16 17:12:20 +02:00
20 changed files with 4363 additions and 66 deletions
+5 -1
View File
@@ -374,7 +374,11 @@ torch = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
torchvision = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
[tool.setuptools.package-data]
lerobot = ["envs/*.json", "annotations/steerable_pipeline/prompts/*.txt"]
lerobot = [
"envs/*.json",
"annotations/steerable_pipeline/prompts/*.txt",
"teleoperators/pico_headset/assets/*.npz",
]
[tool.setuptools.packages.find]
where = ["src"]
+162
View File
@@ -0,0 +1,162 @@
# Unitree G1 — SONIC whole-body control
This package runs NVIDIA's **SONIC** whole-body controller (and the GR00T/Holosoma
locomotion controllers) on the Unitree G1, in MuJoCo simulation or on real hardware.
SONIC turns a high-level movement intent — or a streamed **SMPL** whole-body pose — into
50 Hz joint-position targets. It is a pure-Python/ONNX reimplementation of the SONIC
deploy stack (no `gear_sonic`/torch dependency).
## Controllers
Selected with `--robot.controller=<ClassName>`:
| Controller | Purpose |
| ------------------------------ | ------------------------------------------------------------------------------------------ |
| `SonicWholeBodyController` | SONIC whole-body: locomotion (mode 0), 3-point VR teleop (mode 1), SMPL imitation (mode 2) |
| `GrootLocomotionController` | GR00T locomotion policy |
| `HolosomaLocomotionController` | Holosoma locomotion policy |
On startup the controller **interpolates** from the robot's measured pose into the
policy's commanded target over ~3 s (no snap), and on disconnect (Ctrl-C) it performs a
**graceful damped settle** — holding pose while ramping stiffness to zero over
`--robot.graceful_stop_s` (default 1.5 s) instead of going instantly limp. Both apply in
every mode.
## Requirements
- `onnxruntime` (CPU) **or** `onnxruntime-gpu` (recommended — SONIC runs three ONNX
sessions and is much smoother on GPU). Install the CUDA build that matches your
driver (e.g. `onnxruntime-gpu==1.26.0` for a CUDA-12.x driver). Verify with:
```bash
python -c "import onnxruntime as ort; print(ort.get_available_providers())"
# expect CUDAExecutionProvider in the list for GPU
```
- `mujoco` for simulation (`is_simulation=True`).
- `pyzmq` only if you use the live SMPL stream (pico headset).
- The SONIC ONNX models are downloaded automatically from the `nvidia/GEAR-SONIC` Hub repo.
## Running
**Replay an SMPL dataset (motion imitation):**
```bash
lerobot-replay \
--robot.type=unitree_g1 --robot.controller=SonicWholeBodyController \
--dataset.repo_id=<user>/<smpl_dataset> --dataset.episode=0
```
**Keyboard teleop** (drives locomotion via the native keyboard teleoperator):
```bash
lerobot-teleoperate \
--robot.type=unitree_g1 --robot.controller=SonicWholeBodyController \
--teleop.type=keyboard
```
Controls: `WASD` move · `Q`/`E` turn · `1``8` mode · `9`/`0` speed · `-`/`=` height ·
`R` replan · `Space` emergency-stop.
**PICO headset teleop — SMPL whole-body** (mode 2, needs PICO Motion Trackers):
```bash
# 1) publisher (streams rt/smpl from full-body tracking)
python -m lerobot.teleoperators.pico_headset.pico_publisher --fps 50
# 2) controller
lerobot-teleoperate \
--robot.type=unitree_g1 --robot.controller=SonicWholeBodyController \
--teleop.type=pico_headset
```
**PICO headset teleop — 3-point VR** (mode 1, head + 2 controllers only, **no trackers**):
```bash
# 1) publisher (head + controllers -> 3-point targets + stick locomotion)
python -m lerobot.teleoperators.pico_headset.pico_publisher --fps 50 --headset-source devices
# 2) controller
lerobot-teleoperate \
--robot.type=unitree_g1 --robot.controller=SonicWholeBodyController \
--teleop.type=pico_headset --teleop.mode=vr3
```
3-point controls: left stick move · right stick X turn · right stick Y height ·
`A`+`B` / `X`+`Y` cycle locomotion mode (walk/run/squat/kneel/…) · hands+head track the
upper body. **Calibration**: stand in a neutral rest pose and press `A`+`B`+`X`+`Y` — the
publisher status line flips from `UNCALIBRATED` to `calibrated`. This maps your rest pose
onto the G1's neutral stance and is required before the hands track well; the SMPL
(mode 2) path is self-calibrating and needs no such step.
Both require the XRoboToolkit stack — see below.
## PICO headset / XRoboToolkit install
Live full-body teleop needs the **XRoboToolkit** system (a PC Service on your
workstation + a PICO app on the headset) and its Python binding, `xrobotoolkit_sdk`.
The full hardware + software walkthrough lives in the SONIC repo:
[`docs/source/getting_started/vr_teleop_setup.md`](https://nvlabs.github.io/GR00T-WholeBodyControl/getting_started/vr_teleop_setup.html).
Summary:
1. **PC Service** (workstation) — install and run it before connecting the headset.
- Ubuntu 22.04 / 24.04 (x86_64): prebuilt `.deb` from the
[XRoboToolkit-PC-Service releases](https://github.com/XR-Robotics/XRoboToolkit-PC-Service/releases).
- Jetson (aarch64): the arm64 `.deb`.
- Windows (x64): the Windows PC Service build.
2. **PICO app** — install `XRoboToolkit-PICO-*.apk` on the headset (see the guide),
enable Developer Mode. For **SMPL whole-body** (mode 2) you also need the PICO Motion
Trackers paired/calibrated and "Full body" enabled; for **3-point** (mode 1,
`--headset-source devices`) only Head + Controller + Send are required — no trackers.
3. **`xrobotoolkit_sdk`** — a pybind11/CMake build (not a pip package), from
[`XRoboToolkit-PC-Service-Pybind`](https://github.com/XR-Robotics/XRoboToolkit-PC-Service-Pybind):
- Linux x86_64: `pip install pybind11 cmake` then `bash setup_ubuntu.sh` (or the
SONIC repo's `install_scripts/install_pico.sh`, which builds everything into a
`.venv_teleop`).
- Jetson aarch64: `bash setup_orin.sh` (builds `libPXREARobotSDK.so` from source).
- Windows x64: `pip install pybind11` then `setup_windows.bat` (needs git + an
MSVC/CMake toolchain; uses the prebuilt `PXREARobotSDK.dll`/`.lib`).
4. Connect PICO and workstation to the **same Wi-Fi**, open the XRoboToolkit app, enter
the PC IP, and enable Head/Controller/Send (plus Full-body for SMPL mode 2).
### Platform support
| Platform | Live headset teleop | Notes |
| --------------------------- | ------------------- | ------------------------------------------- |
| Linux x86_64 | ✅ | Guided `install_pico.sh` (SONIC repo) |
| Linux aarch64 (Jetson Orin) | ✅ | `setup_orin.sh` builds the native lib |
| Windows x64 | ✅ (manual) | `setup_windows.bat`; no one-shot env script |
| macOS | ❌ | No PC Service / SDK build for Darwin |
### No hardware required (any platform, incl. macOS/Windows)
The SMPL pipeline can be exercised without a headset or the SDK — the publisher emits
`rt/smpl` frames that the controller consumes exactly as it would from the headset:
```bash
# synthetic motion
python -m lerobot.teleoperators.pico_headset.pico_publisher --fake
# replay a canned SMPL clip
python -m lerobot.teleoperators.pico_headset.pico_publisher --motion-file <clip>.npz
```
## Notes
- SMPL **root motion** into the mode-2 anchor is opt-in (`SonicWholeBodyController(enable_smpl_root=True)`);
it stays off by default (untested on hardware). When enabled, the per-frame root quat is
spherically smoothed (`root_smoothing_alpha`, default 0.15) before it reaches the anchor,
which removes the base-acceleration spikes the raw 30 Hz→50 Hz trajectory used to cause.
- Direct `rt/smpl` subscription without the pico teleoperator is available via
`SonicWholeBodyController(enable_smpl_stream=True, smpl_host=..., smpl_port=...)`.
- 3-point (mode 1) uses the **headset-yaw frame** as its reference and the `A`+`B`+`X`+`Y`
calibration to align to the G1 neutral stance. Calibration maps the operator's rest pose
onto the G1's **standing** (`default_angles`) wrist/neck key-frame poses (position **and**
orientation) computed by FK — the `default_angles` stand-in for gear_sonic's live
measured-q recalibration, since the robot holds `default_angles` at calibration time.
Re-aligning the arms only (preserving the neck level) is available via the calibrator's
`recalibrate_wrists()`.
- 3-point **locomotion** from the PICO sticks follows gear_sonic's `PlannerLoop` exactly:
a yaw accumulator on the right stick and **mode-dependent speed curves** on the left
(slow `0.1+0.5·mag`, run `1.5+3·mag`, walk = planner default). Stick signs replicate
gear_sonic's `get_controller_axes` usage (forward `+ly`, strafe `-lx`, turn `-rx`); since
the publisher forwards the same raw SDK axes, this is the correct convention by construction.
- Startup interpolation and the graceful-stop settle are mode-agnostic; set
`--robot.graceful_stop_s=0` to restore the old instant zero-torque on disconnect.
@@ -65,9 +65,43 @@ class UnitreeG1Config(RobotConfig):
# Cameras (ZMQ-based remote cameras)
cameras: dict[str, CameraConfig] = field(default_factory=dict)
# Synthetic zero-image cameras exposed as ``observation.images.{name}`` (H×W×3
# black frames). Lets image-conditioned policies (e.g. pi0.5 / OpenHLM) run in
# sim before real cameras are wired. Empty = disabled.
empty_cameras: list[str] = field(default_factory=list)
empty_camera_hw: tuple[int, int] = (224, 224)
# Publish Dex3 hand commands (``rt/dex3/{left,right}/cmd``) driven by the OpenHLM
# gripper scalars (``wb.7.pos`` left, ``wb.15.pos`` right). Lets the 43-DoF sim
# (or a real Dex3-equipped G1) show grasping. The scalar in [0, 1] is remapped to
# a curl amount (``hand_open_grip_value`` -> open) and scaled onto
# ``hand_closed_pose`` (7 joints: thumb_0/1/2, middle_0/1, index_0/1). Flip signs
# in ``hand_closed_pose`` if fingers curl the wrong way.
publish_hands: bool = False
hand_open_grip_value: float = 1.0
hand_closed_grip_value: float = 0.0
hand_closed_pose: list[float] = field(
default_factory=lambda: [1.0, 0.9, 0.9, 1.3, 1.3, 1.3, 1.3]
)
hand_kp: float = 1.5
hand_kd: float = 0.1
# Replay recorded camera frames from a LeRobot parquet episode as the camera
# feed (e.g. OpenHLM-data episode). Maps a robot camera name to a parquet image
# column; frames advance one per observation and loop. Lets a VLA see the real
# task video in sim without live cameras. Empty map = disabled.
replay_camera_parquet: str | None = None
replay_camera_map: dict[str, str] = field(default_factory=dict)
replay_camera_loop: bool = True
# Compensates for gravity on the unitree's arms using the arm ik solver
gravity_compensation: bool = False
# Lower-body controller class name, e.g. "GrootLocomotionController" or
# "HolosomaLocomotionController". None disables it.
# Locomotion controller class name, e.g. "GrootLocomotionController",
# "HolosomaLocomotionController", or "SonicWholeBodyController". None disables it.
controller: str | None = None
# On disconnect (e.g. Ctrl-C), seconds to hold the current pose while ramping joint
# stiffness (kp) to zero — a soft, damped settle instead of an instant limp /
# free-fall. 0 disables it (immediate zero-torque). Real robot only.
graceful_stop_s: float = 1.5
@@ -0,0 +1,24 @@
#!/usr/bin/env python
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Unitree G1 locomotion controllers (Groot, Holosoma, SONIC)."""
__all__ = [
"GrootLocomotionController",
"HolosomaLocomotionController",
"SonicWholeBodyController",
"SonicRuntime",
]
@@ -14,20 +14,29 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
import logging
from collections import deque
from typing import TYPE_CHECKING
import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from .g1_utils import (
from lerobot.utils.import_utils import _onnxruntime_available, require_package
from ..g1_utils import (
REMOTE_AXES,
REMOTE_BUTTONS,
G1_29_JointIndex,
get_gravity_orientation,
)
if TYPE_CHECKING or _onnxruntime_available:
import onnxruntime as ort
else:
ort = None
logger = logging.getLogger(__name__)
@@ -83,6 +92,7 @@ class GrootLocomotionController:
control_dt = CONTROL_DT # Expose for unitree_g1.py
def __init__(self):
require_package("onnxruntime", extra="unitree_g1")
# Load policies
self.policy_balance, self.policy_walk = load_groot_policies()
@@ -14,21 +14,34 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
import json
import logging
from typing import TYPE_CHECKING
import numpy as np
import onnx
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from .g1_utils import (
from lerobot.utils.import_utils import _onnx_available, _onnxruntime_available, require_package
from ..g1_utils import (
REMOTE_AXES,
G1_29_JointArmIndex,
G1_29_JointIndex,
get_gravity_orientation,
)
if TYPE_CHECKING or _onnxruntime_available:
import onnxruntime as ort
else:
ort = None
if TYPE_CHECKING or _onnx_available:
import onnx
else:
onnx = None
logger = logging.getLogger(__name__)
DEFAULT_ANGLES = np.zeros(29, dtype=np.float32)
@@ -101,6 +114,8 @@ class HolosomaLocomotionController:
control_dt = CONTROL_DT # Expose for unitree_g1.py
def __init__(self):
require_package("onnxruntime", extra="unitree_g1")
require_package("onnx", extra="unitree_g1")
# Load policy and gains
self.policy, self.kp, self.kd = load_policy()
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,718 @@
#!/usr/bin/env python
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""SONIC full-body controller for Unitree G1."""
from __future__ import annotations
import logging
import math
from collections import deque
from typing import TYPE_CHECKING
import numpy as np
from huggingface_hub import hf_hub_download
from lerobot.teleoperators.pico_headset.smpl_constants import (
LOCO_AXES_PREFIX,
LOCO_BTN_PREFIX,
LOCO_N_AXES,
LOCO_N_BTN,
ROOT_ACTION_DIM,
ROOT_ACTION_PREFIX,
SMPL_ACTION_PREFIX,
SMPL_OBS_DIM as SMPL_ACTION_DIM,
VR3_ORN_DIM,
VR3_ORN_PREFIX,
VR3_POS_DIM,
VR3_POS_PREFIX,
WB_ACTION_DIM,
wb_action_key,
)
from lerobot.utils.import_utils import _onnxruntime_available, require_package
from ..g1_utils import MUJOCO_TO_ISAACLAB, KEYBOARD_KEYS_FIELD, G1_29_JointIndex, lowstate_to_obs
from .sonic_pipeline import (
CONTROL_DT,
DEBUG_PRINT_EVERY,
DEFAULT_ANGLES,
DEFAULT_HEIGHT,
ENCODER_UPDATE_EVERY,
LM,
MOTION_SETS,
MovementState,
PlannerController,
SonicPlanner,
apply_pico_loco_axes,
clamp_mode_params,
compute_kp_kd,
make_ort_session_options,
ort_providers,
process_joystick,
should_replan_request,
snapshot_ms,
)
if TYPE_CHECKING or _onnxruntime_available:
import onnxruntime as ort
else:
ort = None
logger = logging.getLogger(__name__)
# Startup blend duration: over the first control ticks, linearly interpolate every joint
# from the robot's initial measured pose into the policy's commanded target, so control
# eases in without a snap on the first command.
INIT_RAMP_S = 3.0
def _extract_smpl_from_action(action: dict | None) -> np.ndarray | None:
"""Reassemble a (720,) SMPL window from ``smpl.{i}`` action keys, or None.
The pico_headset teleoperator emits the whole-body reference as flat floats so
it flows unchanged through the standard lerobot action pipeline.
"""
# The keys are smpl.0 .. smpl.719; presence of the first element (smpl.0) is the
# sentinel that a full SMPL window was sent this tick. If it's absent, there's no
# whole-body reference, so bail out.
if not action or f"{SMPL_ACTION_PREFIX}0" not in action:
return None
arr = np.fromiter(
(float(action.get(f"{SMPL_ACTION_PREFIX}{i}", 0.0)) for i in range(SMPL_ACTION_DIM)),
dtype=np.float32,
count=SMPL_ACTION_DIM,
)
return arr
def _extract_root_from_action(action: dict | None) -> np.ndarray | None:
"""Reassemble a (4,) SMPL root quaternion (wxyz) from ``root.{i}`` keys, or None."""
if not action or f"{ROOT_ACTION_PREFIX}0" not in action:
return None
q = np.fromiter(
(float(action.get(f"{ROOT_ACTION_PREFIX}{i}", 0.0)) for i in range(ROOT_ACTION_DIM)),
dtype=np.float32,
count=ROOT_ACTION_DIM,
)
n = float(np.linalg.norm(q))
if n < 1e-6:
return None
return q / n
def _extract_vr3_from_action(action: dict | None) -> tuple[np.ndarray, np.ndarray] | None:
"""Reassemble the 3-point VR targets from ``vr3_pos.{i}`` / ``vr3_orn.{i}`` keys.
Returns ``(pos (9,), orn (12,))`` for the [l-wrist, r-wrist, neck] keypoints, or
None when no VR3 reference was sent this tick. Presence of ``vr3_pos.0`` is the
sentinel that a full 3-point frame is available (mirrors the SMPL sentinel).
"""
if not action or f"{VR3_POS_PREFIX}0" not in action:
return None
pos = np.fromiter(
(float(action.get(f"{VR3_POS_PREFIX}{i}", 0.0)) for i in range(VR3_POS_DIM)),
dtype=np.float32,
count=VR3_POS_DIM,
)
orn = np.fromiter(
(float(action.get(f"{VR3_ORN_PREFIX}{i}", 0.0)) for i in range(VR3_ORN_DIM)),
dtype=np.float32,
count=VR3_ORN_DIM,
)
return pos, orn
def _extract_wb34_from_action(action: dict | None) -> np.ndarray | None:
"""Reassemble a dense (34,) whole-body command from ``wb.{i}.pos`` keys, or None.
This is the OpenHLM / pi0.5 joint-based interface: one 34-D vector per tick
(sentinel: presence of ``wb.0.pos``) carrying absolute joint targets in real
units. The ``.pos`` suffix lets these flow through ``lerobot-rollout`` as normal
joint-position action features.
"""
if not action or wb_action_key(0) not in action:
return None
return np.fromiter(
(float(action.get(wb_action_key(i), 0.0)) for i in range(WB_ACTION_DIM)),
dtype=np.float32,
count=WB_ACTION_DIM,
)
def _wb34_to_reference(wb: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
"""Map a 34-D OpenHLM whole-body command to a SONIC mode-0 reference.
Returns ``(ref29, anchor_quat)`` where ``ref29`` is the 29 joint targets in
IsaacLab order (what SONIC's ``motion_joint_positions`` expects) and
``anchor_quat`` (wxyz) encodes the root roll/pitch (yaw=0).
OpenHLM layout : [L-arm 0:7, L-grip 7, R-arm 8:15, R-grip 15,
L-leg 16:22, R-leg 22:28, waist 28:31, root rp+yaw 31:34]
The 29 joints are first assembled in MuJoCo / Unitree-SDK order
([L-leg 0:6, R-leg 6:12, waist 12:15, L-arm 15:22, R-arm 22:29] — the
``G1_29_JointIndex`` grouping OpenHLM uses), then permuted to IsaacLab order via
``MUJOCO_TO_ISAACLAB``. Grippers (7, 15) and yaw-rate (33) are not part of the
29-DoF SONIC reference.
"""
ref_mj = np.zeros(29, np.float32) # MuJoCo / Unitree-SDK grouped order
ref_mj[0:6] = wb[16:22] # left leg
ref_mj[6:12] = wb[22:28] # right leg
ref_mj[12:15] = wb[28:31] # waist
ref_mj[15:22] = wb[0:7] # left arm
ref_mj[22:29] = wb[8:15] # right arm
ref = ref_mj[MUJOCO_TO_ISAACLAB].astype(np.float32) # -> IsaacLab order for SONIC
roll, pitch = float(wb[31]), float(wb[32])
cr, sr, cp, sp = np.cos(roll / 2), np.sin(roll / 2), np.cos(pitch / 2), np.sin(pitch / 2)
anchor = np.array([cr * cp, sr * cp, cr * sp, sr * sp], np.float32) # Rx(roll)·Ry(pitch)
return ref, anchor
def _extract_loco_from_action(action: dict | None) -> tuple[np.ndarray, np.ndarray] | None:
"""Reassemble controller-stick locomotion from ``loco_axes.{i}`` / ``loco_btn.{i}``.
Returns ``(axes (4,) = [lx, ly, rx, ry], buttons (4,) = [A, B, X, Y])`` or None
when no locomotion state was sent this tick (sentinel: ``loco_axes.0``).
"""
if not action or f"{LOCO_AXES_PREFIX}0" not in action:
return None
axes = np.fromiter(
(float(action.get(f"{LOCO_AXES_PREFIX}{i}", 0.0)) for i in range(LOCO_N_AXES)),
dtype=np.float32,
count=LOCO_N_AXES,
)
buttons = np.fromiter(
(float(action.get(f"{LOCO_BTN_PREFIX}{i}", 0.0)) for i in range(LOCO_N_BTN)),
dtype=np.float32,
count=LOCO_N_BTN,
)
return axes, buttons
class SonicRuntime:
"""Shared SONIC control loop state (standalone demo + locomotion controller)."""
def __init__(self, force_cpu: bool = False):
require_package("onnxruntime", extra="unitree_g1")
planner_path = hf_hub_download(repo_id="nvidia/GEAR-SONIC", filename="planner_sonic.onnx")
encoder_path = hf_hub_download(repo_id="nvidia/GEAR-SONIC", filename="model_encoder.onnx")
decoder_path = hf_hub_download(repo_id="nvidia/GEAR-SONIC", filename="model_decoder.onnx")
providers = ort_providers(force_cpu=force_cpu)
self.use_gpu = providers[0] == "CUDAExecutionProvider"
so = make_ort_session_options()
planner_sess = ort.InferenceSession(planner_path, sess_options=so, providers=providers)
encoder_sess = ort.InferenceSession(encoder_path, sess_options=so, providers=providers)
decoder_sess = ort.InferenceSession(decoder_path, sess_options=so, providers=providers)
self.kp, self.kd = compute_kp_kd()
self.ms = MovementState()
self.planner = SonicPlanner(planner_sess, planner_path)
self.controller = PlannerController(self.planner, encoder_sess, decoder_sess)
motion = self.planner.initialize(DEFAULT_ANGLES, self.ms)
self.controller.load_initial_motion(motion)
self.planner.start_subprocess(self.controller, use_gpu=self.use_gpu)
self.step = 0
self.replan_timer = 0.0
self.last_ms = snapshot_ms(self.ms)
@property
def pipeline(self):
return self.controller
def tick(self, obs: dict, *, debug: bool | None = None, use_joystick: bool = True) -> dict:
if not obs:
self.step += 1
return {}
if use_joystick:
process_joystick(obs, self.ms, self.controller)
clamp_mode_params(self.ms)
if self.step > 0:
self.replan_timer += CONTROL_DT
if should_replan_request(self.ms, self.last_ms, self.replan_timer, self.step):
self.planner.request_replan(self.controller.ref_cursor, self.ms)
self.replan_timer = 0.0
self.ms.needs_replan = False
self.last_ms = snapshot_ms(self.ms)
do_enc = self.step % ENCODER_UPDATE_EVERY == 0
if debug is None:
debug = self.step % DEBUG_PRINT_EVERY == 0
action = self.controller.step(obs, update_encoder=do_enc, debug=debug)
result = self.planner.try_get_new_motion()
if result:
self.controller.blend_new_motion(*result)
self.controller.advance_cursor()
self.step += 1
return action
def reset(self):
self.ms = MovementState()
self.controller.reinit_heading = True
self.controller.playing = True
self.step = 0
self.replan_timer = 0.0
self.last_ms = snapshot_ms(self.ms)
def shutdown(self):
self.planner.stop_subprocess()
class SonicWholeBodyController:
"""Full-body SONIC controller for UnitreeG1's background controller thread."""
control_dt = CONTROL_DT
full_body = True
# Advertise a dense 34-D whole-body action space (OpenHLM / pi0.5) so the robot
# exposes ``wb.{i}.pos`` action features and ``lerobot-rollout`` can drive it
# directly with a 34-D VLA policy.
wb_action = True
def __init__(
self,
force_cpu: bool = False,
*,
enable_smpl_root: bool = False,
root_smoothing_alpha: float = 0.15,
enable_smpl_stream: bool = False,
smpl_host: str | None = None,
smpl_port: int | None = None,
):
logger.info("Loading SONIC whole-body controller...")
self._runtime = SonicRuntime(force_cpu=force_cpu)
self.kp = self._runtime.kp
self.kd = self._runtime.kd
self.controller = self._runtime.controller
self.ms = self._runtime.ms
# When True, the per-frame SMPL root quaternion steers the mode-2 anchor.
# Off by default: even with smoothing this changes the anchor/heading and is
# untested on hardware, so it stays opt-in. When enabled, the raw per-frame
# root quat (from a 30 Hz dataset resampled to a 50 Hz loop) is spherically
# smoothed by :meth:`_smooth_root_quat` before it reaches the anchor, which
# removes the root-acceleration spikes (NaN QACC at DOF 0) the unsmoothed
# trajectory caused. ``root_smoothing_alpha`` in (0, 1] is the per-tick blend
# toward the incoming quat (smaller = smoother/laggier, 1 = no smoothing).
self.enable_smpl_root = enable_smpl_root
self._root_smoothing_alpha = float(np.clip(root_smoothing_alpha, 1e-3, 1.0))
self._smoothed_root_quat: np.ndarray | None = None
# Tracks the previous keyboard held-key set so discrete controls (mode,
# motion set, replan, e-stop, WASD direction) fire once per physical press
# instead of every 50 Hz tick while the key is held.
self._prev_keys: set[str] = set()
# Edge state for the PICO A+B / X+Y locomotion-mode cycle (3-point teleop).
self._prev_loco_mode_pair: tuple[bool, bool] = (False, False)
# Startup blend: ease from the robot's initial pose into the first commanded
# policy targets over INIT_RAMP_S (captured on the first control tick).
self._init_ramp_steps = max(1, round(INIT_RAMP_S / CONTROL_DT))
self._init_step = 0
self._start_pose: dict[str, float] = {}
# Tick counter for the dense whole-body (OpenHLM, mode-0) path's encoder cadence.
self._wb_step = 0
# Rolling 50-frame reference trajectory (ref29 + anchor quat) built from the
# stream of per-tick whole-body commands, fed to the encoder as a batch.
self._wb_traj: deque[np.ndarray] = deque(maxlen=50)
self._wb_quat_traj: deque[np.ndarray] = deque(maxlen=50)
# Optional: subscribe directly to the rt/smpl headset stream so full-body
# teleop works with ANY teleoperator (e.g. --teleop.type=unitree_g1 for the
# estop/joystick) before the dedicated pico_headset teleop exists.
self._smpl_host = smpl_host
self._smpl_port = smpl_port
self._smpl_stream = None
if enable_smpl_stream:
self._init_smpl_stream()
logger.info(
"SONIC ready: %s (default mode: %s, smpl_stream=%s)",
MOTION_SETS[0][0],
LM(self.ms.mode).name,
self._smpl_stream is not None,
)
def _init_smpl_stream(self) -> None:
# Lazy import so the zmq dependency is only required when streaming is on.
from lerobot.teleoperators.pico_headset.smpl_stream import (
DEFAULT_SMPL_HOST,
DEFAULT_SMPL_PORT,
SmplStream,
)
host = self._smpl_host or DEFAULT_SMPL_HOST
port = self._smpl_port or DEFAULT_SMPL_PORT
self._smpl_stream = SmplStream(host=host, port=port)
logger.info("SONIC subscribed to rt/smpl @ tcp://%s:%d", host, port)
def _enter_wholebody(self) -> None:
"""Switch into SMPL whole-body tracking (encode_mode 2)."""
self.controller.encode_mode = 2
self.controller.reinit_heading = True
logger.info("SONIC: SMPL stream active -> whole-body tracking (mode 2)")
def _enter_3point(self) -> None:
"""Switch into 3-point VR upper-body teleop (encode_mode 1).
The upper body tracks the VR wrist/neck targets while the lower body /
locomotion keeps running off the planner (joystick/keyboard-driven).
"""
self.controller.encode_mode = 1
self.controller.playing = True
self.controller.reinit_heading = True
self.ms.needs_replan = True
logger.info("SONIC: 3-point VR active -> upper-body tracking + planner locomotion (mode 1)")
def _exit_wholebody(self) -> None:
"""Revert to locomotion/standing (encode_mode 0) after a teleop reference is lost.
Mirrors the 'M' toggle in sonic.py so the handoff is clean: the robot holds
a standing reference and (if a joystick teleop is attached) can be driven.
"""
self.controller.encode_mode = 0
self.controller.playing = True
self.controller.reinit_heading = True
self.ms.needs_replan = True
logger.warning("SONIC: teleop reference lost/stale -> reverting to locomotion (standing)")
def _process_keyboard(self, action: dict | None) -> None:
"""Translate a native KeyboardTeleop's held-key set into MovementState.
Mirrors the standalone SONIC demo's keyboard mapping so locomotion (mode 0/1)
can be driven with ``--teleop.type=keyboard`` instead of the PICO SMPL stream.
Discrete controls act on newly-pressed keys (edge-detected against the previous
tick); inherently-continuous controls (facing turn, height, speed) integrate a
small per-tick delta while the key is held so they feel smooth at 50 Hz.
Controls: WASD move, Q/E turn, 1-8 select mode, 9/0 speed down/up,
-/= height down/up, R replan, Space emergency-stop -> IDLE.
"""
if action is None:
return
keys = action.get(KEYBOARD_KEYS_FIELD)
if keys is None:
return # No KeyboardTeleop attached; leave joystick/SMPL paths untouched.
ms, controller = self.ms, self.controller
held = {k.lower() if isinstance(k, str) and len(k) == 1 else k for k in keys}
prev = self._prev_keys
pressed = held - prev # newly-pressed this tick (edge)
self._prev_keys = held
# ── Discrete: fire once per press ────────────────────────────────────
if "space" in pressed:
ms.mode = LM.IDLE
ms.speed = ms.height = -1.0
ms.has_movement = False
ms.needs_replan = True
controller.playing = False
controller.reinit_heading = True
logger.info("SONIC keyboard: EMERGENCY STOP -> IDLE")
if "r" in pressed:
ms.needs_replan = True
if "n" in pressed or "p" in pressed:
step = 1 if "n" in pressed else -1
ms.motion_set_idx = (ms.motion_set_idx + step) % len(MOTION_SETS)
logger.info("SONIC keyboard: motion set -> %s", MOTION_SETS[ms.motion_set_idx][0])
for digit in ("1", "2", "3", "4", "5", "6", "7", "8"):
if digit in pressed:
idx = int(digit) - 1
modes = MOTION_SETS[ms.motion_set_idx][1]
if 0 <= idx < len(modes):
ms.mode = modes[idx]
ms.has_movement = False
ms.needs_replan = True
controller.playing = True
controller.reinit_heading = True
logger.info("SONIC keyboard: mode -> %s", LM(ms.mode).name)
# WASD sets the movement direction relative to current facing (press to set,
# Space to stop) to match the standalone demo.
if "w" in pressed:
ms.movement_angle = ms.facing_angle
elif "s" in pressed:
ms.movement_angle = ms.facing_angle + math.pi
elif "a" in pressed:
ms.movement_angle = ms.facing_angle + math.pi / 2
elif "d" in pressed:
ms.movement_angle = ms.facing_angle - math.pi / 2
if pressed & {"w", "a", "s", "d"}:
ms.has_movement = True
ms.needs_replan = True
# ── Continuous: integrate a small delta while held ───────────────────
if "q" in held:
ms.facing_angle += 0.02
controller.delta_heading += 0.02
if "e" in held:
ms.facing_angle -= 0.02
controller.delta_heading -= 0.02
if "0" in held:
ms.speed = min(5.0, (ms.speed if ms.speed >= 0 else 1.0) + 0.02)
if "9" in held:
ms.speed = max(0.0, (ms.speed if ms.speed >= 0 else 1.0) - 0.02)
if "=" in held:
ms.height = min(1.0, (ms.height if ms.height >= 0 else DEFAULT_HEIGHT) + 0.005)
if "-" in held:
ms.height = max(0.1, (ms.height if ms.height >= 0 else DEFAULT_HEIGHT) - 0.005)
def _process_pico_loco(self, axes: np.ndarray, buttons: np.ndarray) -> None:
"""Drive locomotion from the PICO controller sticks/buttons (encode_mode 1).
Mirrors gear_sonic's ``PlannerLoop`` VR-3PT tick: left/right sticks steer
movement/facing/speed via :func:`apply_pico_loco_axes` (the faithful gear_sonic
yaw-accumulator + mode-dependent speed curves, not the keyboard-parity map), and
A+B / X+Y edge-cycle the locomotion mode within the current motion set.
"""
lx, ly, rx, ry = (float(v) for v in axes)
apply_pico_loco_axes(lx, ly, rx, ry, self.ms)
# Mode cycling: step linearly through the LocomotionMode enum (A+B = next,
# X+Y = previous), exactly like gear_sonic's PlannerLoop — so the operator can
# reach squat/kneel/crawl, not just the modes in one UI motion set.
a, b, x, y = (v > 0.5 for v in buttons)
ab_now, xy_now = (a and b), (x and y)
ab_prev, xy_prev = self._prev_loco_mode_pair
mode = int(self.ms.mode)
if ab_now and not ab_prev:
mode = min(int(LM.INJURED_WALK), mode + 1)
elif xy_now and not xy_prev:
mode = max(int(LM.IDLE), mode - 1)
if mode != int(self.ms.mode):
self.ms.mode = LM(mode)
self.ms.needs_replan = True
self.controller.playing = True
logger.info("SONIC 3-point: locomotion mode -> %s", LM(self.ms.mode).name)
self._prev_loco_mode_pair = (ab_now, xy_now)
def _run_wholebody34(self, obs: dict, wb: np.ndarray) -> dict:
"""Feed a dense 34-D OpenHLM whole-body command as the mode-0 encoder reference.
The 29 joint targets are held across the encoder lookahead window (zero
velocity) and the root roll/pitch set the anchor orientation, then the
encoder/decoder run directly (planner bypassed). One command per tick, so the
VLA's commanded pose is what SONIC tracks.
"""
ref, anchor = _wb34_to_reference(wb)
c = self.controller
if c.encode_mode != 0:
c.encode_mode = 0
c.reinit_heading = True
# Capture the heading/anchor reference on the first whole-body tick. The
# controller only latches ``init_ref_quat`` (and the base heading) inside
# ``step()`` when ``first_motion or reinit_heading`` — but it already boots in
# mode 0, so the mode-switch guard above misses the very first command and the
# anchor would stay identity. This mirrors the GEAR reference, which seeds
# ``init_ref_quat`` from the first anchor. Must run before the buffers below so
# ``step()`` latches ``motion_body_quats[0]`` = this tick's anchor.
if self._wb_step == 0:
c.reinit_heading = True
# Accumulate the per-tick commands into a rolling 50-frame reference
# trajectory so the encoder's 10-frame, step-5 lookahead sees an actual
# motion sequence (with velocities) instead of one repeated pose. 50 frames
# == chunk horizon == 10 lookahead frames × step 5.
self._wb_traj.append(ref)
self._wb_quat_traj.append(anchor)
traj = np.asarray(self._wb_traj, np.float32) # (L, 29), oldest -> newest
quats = np.asarray(self._wb_quat_traj, np.float32) # (L, 4)
n = len(traj)
# Per-frame velocities from finite differences (rad/s at the control rate).
vel = np.zeros_like(traj)
if n > 1:
vel[1:] = (traj[1:] - traj[:-1]) / CONTROL_DT
vel[0] = vel[1]
with c.motion_lock:
c.motion_joint_positions[:n] = traj
c.motion_joint_velocities[:n] = vel
c.motion_body_quats[:n] = quats
c.motion_body_pos[:n] = 0.0
c.motion_timesteps = n
c.ref_cursor = 0
c.playing = True
do_enc = self._wb_step % ENCODER_UPDATE_EVERY == 0
out = c.step(obs, update_encoder=do_enc, debug=False)
if self._wb_step % 25 == 0:
tgt = np.array([out[f"{m.name}.q"] for m in G1_29_JointIndex], np.float32)
logger.info(
"[WB34] step=%d |ref|mean=%.3f |target|mean=%.3f target_std=%.3f init_ref_quat=%s",
self._wb_step,
float(np.abs(ref).mean()),
float(np.abs(tgt).mean()),
float(tgt.std()),
np.round(c.init_ref_quat, 3).tolist(),
)
self._wb_step += 1
return out
def _smooth_root_quat(self, root_quat: np.ndarray | None) -> np.ndarray | None:
"""Spherically smooth the per-frame SMPL root quaternion (mode-2 anchor).
The reference root trajectory is authored at ~30 Hz and consumed at 50 Hz, so
the raw per-tick quat steps unevenly and injects root-acceleration spikes into
the anchor. This keeps a persistent estimate and shortest-path nlerp-slerps it
toward each incoming (unit) quat by ``root_smoothing_alpha``, yielding a
continuous, rate-matched heading. Quaternions are scalar-first (w, x, y, z).
Returns ``None`` (leaving the anchor self-driven) for an invalid/zero input.
"""
if root_quat is None:
self._smoothed_root_quat = None
return None
q = np.asarray(root_quat, np.float64)
n = np.linalg.norm(q)
if n < 1e-8:
return self._smoothed_root_quat
q = q / n
if self._smoothed_root_quat is None:
self._smoothed_root_quat = q
else:
prev = self._smoothed_root_quat
if np.dot(prev, q) < 0.0: # shortest-path: quats double-cover SO(3)
q = -q
blended = prev + self._root_smoothing_alpha * (q - prev)
self._smoothed_root_quat = blended / (np.linalg.norm(blended) + 1e-12)
return self._smoothed_root_quat.astype(np.float32)
def _startup_blend(self, obs: dict, out: dict) -> dict:
"""Ease into policy control at startup: for the first ``INIT_RAMP_S`` seconds,
interpolate between the robot's pose captured on the first tick and the policy's
live commanded target, so the handoff has no snap.
``out`` is the policy's ``<joint>.q`` target dict for this tick; the blend ratio
climbs 0->1 over the ramp, after which the raw policy target passes through.
"""
if self._init_step >= self._init_ramp_steps or not out:
return out
if self._init_step == 0:
# Capture the robot's actual pose as the interpolation start point.
self._start_pose = {
f"{m.name}.q": float(obs.get(f"{m.name}.q", DEFAULT_ANGLES[m.value]))
for m in G1_29_JointIndex
}
self._init_step += 1
ratio = min(1.0, self._init_step / self._init_ramp_steps)
blended = {
k: self._start_pose.get(k, float(tgt)) * (1.0 - ratio) + float(tgt) * ratio
for k, tgt in out.items()
}
if self._init_step >= self._init_ramp_steps:
logger.info("SONIC startup blend complete -> full policy control")
return blended
def run_step(self, action: dict, lowstate) -> dict:
if lowstate is None:
return {}
obs = lowstate_to_obs(lowstate)
# Keyboard teleop (native KeyboardTeleop) drives the same locomotion intent
# the joystick does; applied before the SMPL check so whole-body tracking
# still takes priority when a headset stream is present.
self._process_keyboard(action)
# Prefer SMPL delivered via the teleop action (pico_headset). Fall back to a
# direct rt/smpl subscription when enabled (enable_smpl_stream). A stale
# stream (headset silent past its timeout) is treated as "no SMPL" so the
# robot doesn't stay frozen tracking the last pose.
# Dense whole-body command (OpenHLM / pi0.5 joint interface) takes priority:
# a single 34-D vector drives the mode-0 joint reference directly.
wb = _extract_wb34_from_action(action)
if wb is not None:
return self._startup_blend(obs, self._run_wholebody34(obs, wb))
self._wb_miss = getattr(self, "_wb_miss", 0) + 1
if self._wb_miss % 50 == 1:
akeys = [k for k in action if isinstance(k, str)]
logger.info(
"[WB34] no wb.*.pos in action this tick (miss=%d). action keys sample: %s",
self._wb_miss,
akeys[:8],
)
smpl = _extract_smpl_from_action(action)
root_quat = _extract_root_from_action(action)
vr3 = _extract_vr3_from_action(action)
loco = _extract_loco_from_action(action)
if smpl is None and vr3 is None and self._smpl_stream is not None:
window = self._smpl_stream.step()
if self._smpl_stream.has_data and not self._smpl_stream.is_stale:
smpl = window
root_quat = np.asarray(self._smpl_stream.root_quat, np.float32)
# VR3 is independent of the SMPL window: the controller-state source
# (head + controllers only) sends 3-point targets with no SMPL frame.
elif self._smpl_stream.has_fresh_vr3:
vr3 = (self._smpl_stream.vr3_pos, self._smpl_stream.vr3_orn)
if self._smpl_stream.has_fresh_loco:
loco = (self._smpl_stream.loco_axes, self._smpl_stream.loco_buttons)
if smpl is not None:
# Full-body whole-body tracking: SMPL drives the reference, not joystick.
if self.controller.encode_mode != 2:
self._enter_wholebody()
self.controller.smpl_joints_10frame_step1 = smpl
# Root orientation steers the mode-2 anchor/heading, but only when
# explicitly enabled (see enable_smpl_root); the raw per-frame quat is
# spherically smoothed first so the 30->50 Hz resample doesn't spike the
# anchor. Disabled -> anchor stays self-driven.
self.controller.smpl_root_quat = (
self._smooth_root_quat(root_quat) if self.enable_smpl_root else None
)
out = self._runtime.tick(obs, debug=False, use_joystick=False)
elif vr3 is not None:
# 3-point VR teleop: upper body tracks the wrist/neck targets; the lower
# body / locomotion keeps running off the planner, so the joystick (and
# keyboard) still steer walking/turning underneath.
if self.controller.encode_mode != 1:
self._enter_3point()
self.controller.vr_3point_local_target = vr3[0]
self.controller.vr_3point_local_orn_target = vr3[1]
# Replicate the original encode_mode-1 handling: when the PICO controller
# sticks are forwarded, drive locomotion from them directly (and skip the
# wireless-remote joystick read). Otherwise leave the remote/keyboard path.
if loco is not None:
self._process_pico_loco(loco[0], loco[1])
out = self._runtime.tick(obs, debug=False, use_joystick=False)
else:
out = self._runtime.tick(obs, debug=False, use_joystick=True)
else:
# No (or stale) teleop reference: fall back to locomotion so the robot stays balanced.
if self.controller.encode_mode != 0:
self.controller.smpl_root_quat = None
self._smoothed_root_quat = None
self._exit_wholebody()
out = self._runtime.tick(obs, debug=False)
# Startup interpolation: blend from the robot's initial pose into the policy's
# commanded target over INIT_RAMP_S, regardless of mode.
return self._startup_blend(obs, out)
def reset(self):
self._runtime.reset()
self._init_step = 0 # re-run the startup blend after a reset
self._start_pose = {}
self._smoothed_root_quat = None
self._wb_step = 0
self._wb_traj.clear()
self._wb_quat_traj.clear()
def shutdown(self):
if self._smpl_stream is not None:
self._smpl_stream.close()
self._runtime.shutdown()
+173 -2
View File
@@ -23,10 +23,102 @@ import numpy as np
NUM_MOTORS = 29
# Joint-order permutations between the two 29-DoF layouts used across the G1 stack:
# IsaacLab (policy/training order) and MuJoCo (deploy order). ``a[ISAACLAB_TO_MUJOCO]``
# reorders an IsaacLab-ordered vector into MuJoCo order, and vice-versa.
ISAACLAB_TO_MUJOCO = np.array(
[
0,
3,
6,
9,
13,
17,
1,
4,
7,
10,
14,
18,
2,
5,
8,
11,
15,
19,
21,
23,
25,
27,
12,
16,
20,
22,
24,
26,
28,
],
dtype=np.int32,
)
MUJOCO_TO_ISAACLAB = np.array(
[
0,
6,
12,
1,
7,
13,
2,
8,
14,
3,
9,
15,
22,
4,
10,
16,
23,
5,
11,
17,
24,
18,
25,
19,
26,
20,
27,
21,
28,
],
dtype=np.int32,
)
REMOTE_AXES = ("remote.lx", "remote.ly", "remote.rx", "remote.ry")
REMOTE_BUTTONS = tuple(f"remote.button.{i}" for i in range(16))
REMOTE_KEYS = REMOTE_AXES + REMOTE_BUTTONS
# Reserved action-dict field used to forward the set of currently-pressed keyboard
# keys from a KeyboardTeleop through the standard action pipeline to the SONIC
# whole-body controller (see SonicWholeBodyController._process_keyboard).
KEYBOARD_KEYS_FIELD = "keyboard.keys"
# ── Dense whole-body joint reference (SONIC encode_mode 0, OpenHLM / pi0.5) ──────
# A single 34-D whole-body command per tick, in the OpenHLM action layout:
# [L-arm(7), L-grip(1), R-arm(7), R-grip(1), L-leg(6), R-leg(6), waist(3),
# root roll/pitch + yaw-rate(3)]
# Fed as flat scalars ``wb.0.pos .. wb.33.pos``. The ``.pos`` suffix makes these
# behave like ordinary joint-position action features so ``lerobot-rollout`` routes
# them straight from a 34-D VLA (OpenHLM / pi0.5) onto the robot.
WB_ACTION_PREFIX = "wb."
WB_ACTION_DIM = 34
def wb_action_key(i: int) -> str:
"""Action-dict key for the ``i``-th whole-body command scalar (``wb.{i}.pos``)."""
return f"{WB_ACTION_PREFIX}{i}.pos"
def default_remote_input() -> dict[str, float]:
"""Return a zeroed-out remote input dict (axes + buttons)."""
@@ -63,13 +155,92 @@ class G1_29_JointArmIndex(IntEnum):
kRightWristYaw = 28
def lowstate_to_obs(lowstate) -> dict:
"""Build a robot observation dict from a Unitree lowstate.
Shared by ``UnitreeG1.get_observation`` and the SONIC pipeline so the
lowstate -> obs mapping lives in exactly one place. Keys match the
``<joint>.q``/``imu.*`` schema consumed across the controllers.
"""
obs: dict = {}
for motor in G1_29_JointIndex:
idx = motor.value
obs[f"{motor.name}.q"] = lowstate.motor_state[idx].q
obs[f"{motor.name}.dq"] = lowstate.motor_state[idx].dq
obs[f"{motor.name}.tau"] = lowstate.motor_state[idx].tau_est
imu = lowstate.imu_state
if imu.gyroscope:
obs["imu.gyro.x"] = imu.gyroscope[0]
obs["imu.gyro.y"] = imu.gyroscope[1]
obs["imu.gyro.z"] = imu.gyroscope[2]
if imu.accelerometer:
obs["imu.accel.x"] = imu.accelerometer[0]
obs["imu.accel.y"] = imu.accelerometer[1]
obs["imu.accel.z"] = imu.accelerometer[2]
if imu.quaternion:
obs["imu.quat.w"] = imu.quaternion[0]
obs["imu.quat.x"] = imu.quaternion[1]
obs["imu.quat.y"] = imu.quaternion[2]
obs["imu.quat.z"] = imu.quaternion[3]
if imu.rpy:
obs["imu.rpy.roll"] = imu.rpy[0]
obs["imu.rpy.pitch"] = imu.rpy[1]
obs["imu.rpy.yaw"] = imu.rpy[2]
wr = getattr(lowstate, "wireless_remote", None)
if wr:
obs["wireless_remote"] = bytes(wr) if not isinstance(wr, (bytes, bytearray)) else wr
return obs
def obs_to_wb34_state(obs: dict) -> np.ndarray:
"""Build the 34-D OpenHLM / pi0.5 proprio state from a G1 observation dict.
Mirrors the whole-body *action* layout so the policy sees state and action in
the same coordinates::
[L-arm(7), L-grip(1), R-arm(7), R-grip(1),
L-leg(6), R-leg(6), waist(3), root roll/pitch + yaw-rate(3)]
Joint positions come from the ``<joint>.q`` obs keys, which are already in
MuJoCo / Unitree-SDK order — the same body-part grouping OpenHLM uses
([L-leg 0:6, R-leg 6:12, waist 12:15, L-arm 15:22, R-arm 22:29]) — so they are
regrouped directly (no IsaacLab permutation). The G1 has no grippers in its
29-DoF body, so both gripper slots are 0. Root roll/pitch are the IMU RPY and
the last slot is the IMU yaw rate (gyro z).
"""
q_mj = np.array(
[float(obs.get(f"{m.name}.q", 0.0)) for m in G1_29_JointIndex],
dtype=np.float32,
)
lleg, rleg, waist = q_mj[0:6], q_mj[6:12], q_mj[12:15]
larm, rarm = q_mj[15:22], q_mj[22:29]
state = np.zeros(34, dtype=np.float32)
state[0:7] = larm
# state[7] left gripper — none on 29-DoF G1
state[8:15] = rarm
# state[15] right gripper — none on 29-DoF G1
state[16:22] = lleg
state[22:28] = rleg
state[28:31] = waist
state[31] = float(obs.get("imu.rpy.roll", 0.0))
state[32] = float(obs.get("imu.rpy.pitch", 0.0))
state[33] = float(obs.get("imu.gyro.z", 0.0))
return state
def make_locomotion_controller(name: str | None):
"""Instantiate a locomotion controller by class name. Returns None if name is None."""
if name is None:
return None
controllers = {
"GrootLocomotionController": "lerobot.robots.unitree_g1.gr00t_locomotion",
"HolosomaLocomotionController": "lerobot.robots.unitree_g1.holosoma_locomotion",
"GrootLocomotionController": "lerobot.robots.unitree_g1.controllers.gr00t_locomotion",
"HolosomaLocomotionController": "lerobot.robots.unitree_g1.controllers.holosoma_locomotion",
"SonicWholeBodyController": "lerobot.robots.unitree_g1.controllers.sonic_whole_body",
}
module_path = controllers.get(name)
if module_path is None:
+226 -54
View File
@@ -33,12 +33,14 @@ from ..robot import Robot
from .config_unitree_g1 import UnitreeG1Config
from .g1_kinematics import G1_29_ArmIK
from .g1_utils import (
KEYBOARD_KEYS_FIELD,
REMOTE_AXES,
REMOTE_KEYS,
G1_29_JointArmIndex,
G1_29_JointIndex,
default_remote_input,
lowstate_to_obs,
make_locomotion_controller,
obs_to_wb34_state,
)
if TYPE_CHECKING or _unitree_sdk_available:
@@ -47,8 +49,12 @@ if TYPE_CHECKING or _unitree_sdk_available:
ChannelPublisher as _SDKChannelPublisher,
ChannelSubscriber as _SDKChannelSubscriber,
)
from unitree_sdk2py.idl.default import unitree_hg_msg_dds__LowCmd_
from unitree_sdk2py.idl.default import (
unitree_hg_msg_dds__HandCmd_ as hg_HandCmd_default,
unitree_hg_msg_dds__LowCmd_,
)
from unitree_sdk2py.idl.unitree_hg.msg.dds_ import (
HandCmd_ as hg_HandCmd,
LowCmd_ as hg_LowCmd,
LowState_ as hg_LowState,
)
@@ -58,6 +64,8 @@ else:
_SDKChannelPublisher = None
_SDKChannelSubscriber = None
unitree_hg_msg_dds__LowCmd_ = None
hg_HandCmd_default = None
hg_HandCmd = None
hg_LowCmd = None
hg_LowState = None
CRC = None
@@ -157,6 +165,37 @@ class UnitreeG1(Robot):
self.controller_input = default_remote_input()
self.controller_output = {}
# Replay-camera state (decoded frames per robot camera name + play cursor).
self._replay_frames: dict[str, list[np.ndarray]] = {}
self._replay_len = 0
self._replay_idx = 0
if config.replay_camera_parquet and config.replay_camera_map:
self._load_replay_frames()
def _load_replay_frames(self) -> None:
"""Decode recorded episode frames from a parquet into per-camera image lists."""
import io
import pyarrow.parquet as pq
from PIL import Image
table = pq.read_table(self.config.replay_camera_parquet)
cols = {col: table.column(col).to_pylist() for col in self.config.replay_camera_map.values()}
self._replay_len = table.num_rows
def decode(cell) -> np.ndarray:
data = cell["bytes"] if isinstance(cell, dict) else cell
return np.asarray(Image.open(io.BytesIO(data)).convert("RGB"), dtype=np.uint8)
for cam_name, column in self.config.replay_camera_map.items():
self._replay_frames[cam_name] = [decode(c) for c in cols[column]]
logger.info(
"Loaded %d replay frames for cameras %s from %s",
self._replay_len,
list(self.config.replay_camera_map),
self.config.replay_camera_parquet,
)
def _subscribe_lowstate(self): # polls robot state @ 250Hz
while not self._shutdown_event.is_set():
start_time = time.time()
@@ -231,15 +270,54 @@ class UnitreeG1(Robot):
features[f"{cam}_depth"] = (cfg.height, cfg.width, 1)
return features
@property
def _wb_state_ft(self) -> dict[str, type]:
"""34-D whole-body proprio state (``wb_state.{i}.pos``) for dense controllers.
Exposed only when the controller consumes a dense whole-body command
(OpenHLM / pi0.5). These ``.pos`` scalars are aggregated by the rollout
pipeline into a single 34-D ``observation.state`` for the policy.
"""
if not getattr(self.controller, "wb_action", False):
return {}
from .g1_utils import WB_ACTION_DIM
return {f"wb_state.{i}.pos": float for i in range(WB_ACTION_DIM)}
@property
def _empty_cameras_ft(self) -> dict[str, tuple]:
"""Synthetic zero-image cameras (see ``UnitreeG1Config.empty_cameras``)."""
h, w = self.config.empty_camera_hw
return {name: (h, w, 3) for name in self.config.empty_cameras}
@property
def _replay_cameras_ft(self) -> dict[str, tuple]:
"""Replay cameras, shaped from their first decoded frame."""
return {name: frames[0].shape for name, frames in self._replay_frames.items() if frames}
@cached_property
def observation_features(self) -> dict[str, type | tuple]:
return {**self._motors_ft, **self._cameras_ft}
return {
**self._motors_ft,
**self._wb_state_ft,
**self._empty_cameras_ft,
**self._replay_cameras_ft,
**self._cameras_ft,
}
@cached_property
def action_features(self) -> dict[str, type]:
if self.controller is None:
return {f"{G1_29_JointIndex(motor).name}.q": float for motor in G1_29_JointIndex}
# Dense whole-body controllers (SONIC / OpenHLM, pi0.5) consume a single
# 34-D command per tick. Expose it as ``wb.{i}.pos`` joint-position features
# so ``lerobot-rollout`` maps a 34-D policy output straight onto the robot.
if getattr(self.controller, "wb_action", False):
from .g1_utils import WB_ACTION_DIM, wb_action_key
return {wb_action_key(i): float for i in range(WB_ACTION_DIM)}
arm_features = {f"{G1_29_JointArmIndex(motor).name}.q": float for motor in G1_29_JointArmIndex}
remote_features = dict.fromkeys(REMOTE_AXES, float)
return {**arm_features, **remote_features}
@@ -311,6 +389,17 @@ class UnitreeG1(Robot):
self.lowstate_subscriber = self._ChannelSubscriber(kTopicLowState, hg_LowState)
self.lowstate_subscriber.Init()
# Dex3 hand command publishers (grasping). Driven by the OpenHLM grip scalars.
self._hand_publishers = {}
if self.config.publish_hands:
self._left_hand_cmd = hg_HandCmd_default()
self._right_hand_cmd = hg_HandCmd_default()
self._hand_publishers["left"] = self._ChannelPublisher("rt/dex3/left/cmd", hg_HandCmd)
self._hand_publishers["right"] = self._ChannelPublisher("rt/dex3/right/cmd", hg_HandCmd)
for pub in self._hand_publishers.values():
pub.Init()
logger.info("Dex3 hand command publishers initialized (rt/dex3/{left,right}/cmd)")
# Start subscribe thread to read robot state
self.subscribe_thread = threading.Thread(target=self._subscribe_lowstate)
self.subscribe_thread.start()
@@ -343,6 +432,9 @@ class UnitreeG1(Robot):
self.kp = np.array(self.config.kp, dtype=np.float32)
self.kd = np.array(self.config.kd, dtype=np.float32)
if self.controller is not None and hasattr(self.controller, "kp"):
self.kp = np.array(self.controller.kp, dtype=np.float32)
self.kd = np.array(self.controller.kd, dtype=np.float32)
for joint in G1_29_JointIndex:
self.msg.motor_cmd[joint].mode = 1
@@ -371,13 +463,50 @@ class UnitreeG1(Robot):
except Exception as e:
logger.warning(f"Failed to send zero-torque on disconnect: {e}")
def disconnect(self):
# Put robot in passive mode before stopping threads
if not self.config.is_simulation:
self._send_zero_torque()
def _graceful_stop(self) -> None:
"""Soft shutdown: hold the current pose and ramp joint stiffness (kp) to zero
over ``graceful_stop_s`` while keeping damping (kd), then go passive.
# Signal thread to stop and unblock any waits
Prevents the robot from collapsing the instant control ends (a bare
zero-torque command is kp=kd=0 ≈ free-fall). Must run after the controller
loop has stopped so the two aren't publishing at once.
"""
if self.config.graceful_stop_s <= 0:
self._send_zero_torque()
return
with self._lowstate_lock:
lowstate = self._lowstate
if lowstate is None:
self._send_zero_torque()
return
q_hold = {f"{motor.name}.q": lowstate.motor_state[motor.value].q for motor in G1_29_JointIndex}
kp = np.array(self.kp, dtype=np.float32)
kd = np.array(self.kd, dtype=np.float32)
zeros = np.zeros(29, dtype=np.float32)
dt = self.controller.control_dt if self.controller is not None else self.config.control_dt
steps = max(1, int(self.config.graceful_stop_s / dt))
logger.info("Graceful stop: damping down over %.1fs", self.config.graceful_stop_s)
for i in range(steps):
ratio = (i + 1) / steps
self.publish_lowcmd(q_hold, kp=kp * (1.0 - ratio), kd=kd, tau=zeros)
time.sleep(dt)
self._send_zero_torque()
def disconnect(self):
# Stop the controller loop first so it isn't fighting the shutdown ramp.
self._shutdown_event.set()
if self._controller_thread is not None:
self._controller_thread.join(timeout=2.0)
if self._controller_thread.is_alive():
logger.warning("Controller thread did not stop cleanly")
# Soft, damped settle instead of an instant limp (real robot only; the
# subscribe thread is still alive here to supply the current pose).
if not self.config.is_simulation:
self._graceful_stop()
if self.controller is not None and hasattr(self.controller, "shutdown"):
self.controller.shutdown()
# Wait for subscribe thread to finish
if self.subscribe_thread is not None:
@@ -385,12 +514,6 @@ class UnitreeG1(Robot):
if self.subscribe_thread.is_alive():
logger.warning("Subscribe thread did not stop cleanly")
# Wait for controller thread to finish
if self._controller_thread is not None:
self._controller_thread.join(timeout=2.0)
if self._controller_thread.is_alive():
logger.warning("Controller thread did not stop cleanly")
# Close simulation environment
if self.config.is_simulation and self.sim_env is not None:
try:
@@ -422,44 +545,33 @@ class UnitreeG1(Robot):
if lowstate is None:
return {}
obs = {}
# Motors + IMU + wireless remote (shared lowstate -> obs mapping)
obs = lowstate_to_obs(lowstate)
# Motors - q, dq, tau for all joints
for motor in G1_29_JointIndex:
name = motor.name
idx = motor.value
obs[f"{name}.q"] = lowstate.motor_state[idx].q
obs[f"{name}.dq"] = lowstate.motor_state[idx].dq
obs[f"{name}.tau"] = lowstate.motor_state[idx].tau_est
# Dense whole-body controllers (OpenHLM / pi0.5): expose the 34-D proprio
# state as ``wb_state.{i}.pos`` so the rollout aggregates it into
# ``observation.state`` for the policy.
if getattr(self.controller, "wb_action", False):
wb_state = obs_to_wb34_state(obs)
for i, v in enumerate(wb_state):
obs[f"wb_state.{i}.pos"] = float(v)
# IMU - gyroscope
if lowstate.imu_state.gyroscope:
obs["imu.gyro.x"] = lowstate.imu_state.gyroscope[0]
obs["imu.gyro.y"] = lowstate.imu_state.gyroscope[1]
obs["imu.gyro.z"] = lowstate.imu_state.gyroscope[2]
# Synthetic empty cameras: black frames so image-conditioned policies run
# before real cameras are wired.
if self.config.empty_cameras:
h, w = self.config.empty_camera_hw
black = np.zeros((h, w, 3), dtype=np.uint8)
for name in self.config.empty_cameras:
obs[name] = black
# IMU - accelerometer
if lowstate.imu_state.accelerometer:
obs["imu.accel.x"] = lowstate.imu_state.accelerometer[0]
obs["imu.accel.y"] = lowstate.imu_state.accelerometer[1]
obs["imu.accel.z"] = lowstate.imu_state.accelerometer[2]
# IMU - quaternion
if lowstate.imu_state.quaternion:
obs["imu.quat.w"] = lowstate.imu_state.quaternion[0]
obs["imu.quat.x"] = lowstate.imu_state.quaternion[1]
obs["imu.quat.y"] = lowstate.imu_state.quaternion[2]
obs["imu.quat.z"] = lowstate.imu_state.quaternion[3]
# IMU - rpy
if lowstate.imu_state.rpy:
obs["imu.rpy.roll"] = lowstate.imu_state.rpy[0]
obs["imu.rpy.pitch"] = lowstate.imu_state.rpy[1]
obs["imu.rpy.yaw"] = lowstate.imu_state.rpy[2]
# Wireless remote (raw bytes for teleoperator)
if lowstate.wireless_remote:
obs["wireless_remote"] = lowstate.wireless_remote
# Replay cameras: serve the current recorded frame per camera, then advance.
if self._replay_len:
idx = self._replay_idx
if idx >= self._replay_len:
idx = self._replay_len - 1 if not self.config.replay_camera_loop else idx % self._replay_len
for name, frames in self._replay_frames.items():
obs[name] = frames[idx]
self._replay_idx += 1
# Cameras - read images from ZMQ cameras
for cam_name, cam in self._cameras.items():
@@ -473,9 +585,13 @@ class UnitreeG1(Robot):
def send_action(self, action: RobotAction) -> RobotAction:
action_to_publish = action
if self.controller is not None:
self._update_controller_action(action)
if self.config.publish_hands and getattr(self.controller, "wb_action", False):
self._publish_hand_cmds(action)
if getattr(self.controller, "full_body", False):
return action
# Controller thread owns legs/waist. Here we only update joystick inputs
# and publish arm targets from the teleoperator.
self._update_controller_action(action)
arm_prefixes = tuple(j.name for j in G1_29_JointArmIndex)
action_to_publish = {
key: value
@@ -503,11 +619,67 @@ class UnitreeG1(Robot):
return action
def _update_controller_action(self, action: RobotAction) -> None:
"""Update controller input state from incoming teleop action."""
"""Update controller input state from an incoming teleop action.
Controller-agnostic: every value-carrying key is forwarded verbatim into
``controller_input`` (whole-body ``wb.{i}.pos`` from a 34-D VLA, or whatever a
future controller expects), and each controller extracts only the keys it
understands. The robot deliberately does not enumerate any controller's key
schema here.
KeyboardTeleop is the one special case: it emits the currently-pressed keys as
bare action keys with a ``None`` value (``dict.fromkeys(pressed, None)``), so
those are collected into a single held-key set under ``KEYBOARD_KEYS_FIELD``,
rebuilt each tick so releases clear. Special keys arrive as pynput objects and
are normalised to their name ("space", ...).
"""
with self._controller_action_lock:
for key in REMOTE_KEYS:
if key in action:
self.controller_input[key] = action[key]
self.controller_input[KEYBOARD_KEYS_FIELD] = {
(k if isinstance(k, str) else getattr(k, "name", str(k)))
for k, value in action.items()
if value is None
}
for key, value in action.items():
if isinstance(key, str) and value is not None:
self.controller_input[key] = value
def _publish_hand_cmds(self, action: RobotAction) -> None:
"""Drive the Dex3 hands from the OpenHLM grip scalars in a 34-D wb action.
``wb.7.pos`` is the left grip and ``wb.15.pos`` the right grip. Each scalar in
[0, 1] (``hand_open_grip_value`` == fully open) is turned into a curl amount and
scaled onto ``hand_closed_pose`` (7 joints), then published as a PD target on
``rt/dex3/{left,right}/cmd`` so the fingers close when the policy grips.
"""
if not self._hand_publishers:
return
from .g1_utils import wb_action_key
open_val = float(self.config.hand_open_grip_value)
closed_val = float(self.config.hand_closed_grip_value)
closed_pose = self.config.hand_closed_pose
kp, kd = float(self.config.hand_kp), float(self.config.hand_kd)
span = (closed_val - open_val) or 1.0
def curl_amount(grip: float) -> float:
# Fraction of the way from the open scalar to the closed scalar, in [0, 1].
return float(min(max((grip - open_val) / span, 0.0), 1.0))
for side, grip_idx, cmd in (
("left", 7, self._left_hand_cmd),
("right", 15, self._right_hand_cmd),
):
grip = action.get(wb_action_key(grip_idx))
if grip is None:
continue
amount = curl_amount(float(grip))
for i, closed_q in enumerate(closed_pose):
cmd.motor_cmd[i].q = float(closed_q) * amount
cmd.motor_cmd[i].dq = 0.0
cmd.motor_cmd[i].kp = kp
cmd.motor_cmd[i].kd = kd
cmd.motor_cmd[i].tau = 0.0
self._hand_publishers[side].Write(cmd)
@property
def is_calibrated(self) -> bool:
@@ -116,6 +116,7 @@ from lerobot.teleoperators import ( # noqa: F401
omx_leader,
openarm_leader,
openarm_mini,
pico_headset,
reachy2_teleoperator,
rebot_102_leader,
so_leader,
@@ -0,0 +1,20 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .config_pico_headset import PicoHeadsetConfig
from .pico_headset import PicoHeadset
__all__ = ["PicoHeadset", "PicoHeadsetConfig"]
@@ -0,0 +1,41 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from ..config import TeleoperatorConfig
@TeleoperatorConfig.register_subclass("pico_headset")
@dataclass
class PicoHeadsetConfig(TeleoperatorConfig):
"""PICO full-body headset teleop: live SMPL over the rt/smpl ZMQ stream.
Consumes the ``rt/smpl`` channel published by the GEAR PICO manager
(``gear_sonic/scripts/pico_manager_thread_server.py``) and emits the whole-body
SONIC reference window (``encode_mode == 2``) for SonicWholeBodyController.
"""
smpl_host: str = "127.0.0.1"
"""Host of the pico_manager rt/smpl publisher (the laptop bridging the PICO)."""
smpl_port: int = 5560
"""Port of the rt/smpl publisher."""
stale_after_s: float = 0.5
"""Warn if no fresh headset frame arrives within this many seconds."""
mode: str = "smpl"
"""Teleop reference to emit: ``"smpl"`` for whole-body imitation (SONIC
encode_mode 2) or ``"vr3"`` for sparse 3-point upper-body teleop (encode_mode 1,
lower body driven by the joystick/keyboard planner)."""
@@ -0,0 +1,148 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PICO full-body headset teleoperator (live SMPL -> SONIC whole-body)."""
import logging
from typing import Any
from lerobot.types import RobotAction
from ..teleoperator import Teleoperator
from .config_pico_headset import PicoHeadsetConfig
from .smpl_constants import (
LOCO_AXES_PREFIX,
LOCO_BTN_PREFIX,
LOCO_N_AXES,
LOCO_N_BTN,
ROOT_ACTION_DIM,
ROOT_ACTION_PREFIX,
SMPL_ACTION_PREFIX,
SMPL_OBS_DIM,
VR3_ORN_DIM,
VR3_ORN_PREFIX,
VR3_POS_DIM,
VR3_POS_PREFIX,
)
from .smpl_stream import SmplStream
logger = logging.getLogger(__name__)
class PicoHeadset(Teleoperator):
"""Streams full-body SMPL from a PICO headset as a SONIC whole-body reference.
Subscribes to the ``rt/smpl`` ZMQ channel and, once real frames are flowing,
emits the 720-element encoder window as ``smpl.{i}`` floats. Before the first
frame arrives it emits no SMPL keys, so the robot stays in safe locomotion mode
rather than tracking a zero pose.
"""
config_class = PicoHeadsetConfig
name = "pico_headset"
def __init__(self, config: PicoHeadsetConfig):
super().__init__(config)
self.config = config
self._stream: SmplStream | None = None
@property
def action_features(self) -> dict:
if self.config.mode == "vr3":
feats = {f"{VR3_POS_PREFIX}{i}": float for i in range(VR3_POS_DIM)}
feats.update({f"{VR3_ORN_PREFIX}{i}": float for i in range(VR3_ORN_DIM)})
# Controller-stick locomotion travels alongside the VR targets.
feats.update({f"{LOCO_AXES_PREFIX}{i}": float for i in range(LOCO_N_AXES)})
feats.update({f"{LOCO_BTN_PREFIX}{i}": float for i in range(LOCO_N_BTN)})
return feats
feats = {f"{SMPL_ACTION_PREFIX}{i}": float for i in range(SMPL_OBS_DIM)}
feats.update({f"{ROOT_ACTION_PREFIX}{i}": float for i in range(ROOT_ACTION_DIM)})
return feats
@property
def feedback_features(self) -> dict:
return {}
@property
def is_connected(self) -> bool:
return self._stream is not None
def connect(self, calibrate: bool = True) -> None:
if self._stream is not None:
raise RuntimeError(f"{self} already connected")
self._stream = SmplStream(
host=self.config.smpl_host,
port=self.config.smpl_port,
stale_after_s=self.config.stale_after_s,
)
logger.info(
"PicoHeadset subscribed to rt/smpl @ tcp://%s:%d",
self.config.smpl_host,
self.config.smpl_port,
)
@property
def is_calibrated(self) -> bool:
return True
def calibrate(self) -> None:
# Calibration happens on the headset / pico_manager side, not here.
pass
def configure(self) -> None:
pass
def get_action(self) -> RobotAction:
if self._stream is None:
raise RuntimeError(f"{self} is not connected")
window = self._stream.step()
# Emit a reference only while the headset is actively streaming: hold back
# before the first frame (so the controller doesn't track an all-zero
# collapsed pose) and once the stream goes stale (so the controller falls
# back to a safe standing/locomotion mode instead of freezing on the last
# pose).
if self.config.mode == "vr3":
# Sparse 3-point upper-body teleop (encode_mode 1). Gated on fresh vr3_*
# frames only (independent of the SMPL window), so the controller-state
# source (head + controllers, no body tracking) works. Emit nothing
# otherwise and stay in locomotion.
if not self._stream.has_fresh_vr3:
return {}
action = {f"{VR3_POS_PREFIX}{i}": float(v) for i, v in enumerate(self._stream.vr3_pos)}
action.update({f"{VR3_ORN_PREFIX}{i}": float(v) for i, v in enumerate(self._stream.vr3_orn)})
# Forward controller-stick locomotion when present, so the planner can
# steer walking/turning under the upper-body tracking (encode_mode 1).
if self._stream.has_fresh_loco:
action.update(
{f"{LOCO_AXES_PREFIX}{i}": float(v) for i, v in enumerate(self._stream.loco_axes)}
)
action.update(
{f"{LOCO_BTN_PREFIX}{i}": float(v) for i, v in enumerate(self._stream.loco_buttons)}
)
return action
if not self._stream.has_data or self._stream.is_stale:
return {}
action = {f"{SMPL_ACTION_PREFIX}{i}": float(v) for i, v in enumerate(window)}
action.update({f"{ROOT_ACTION_PREFIX}{i}": float(v) for i, v in enumerate(self._stream.root_quat)})
return action
def send_feedback(self, feedback: dict[str, Any]) -> None:
pass
def disconnect(self) -> None:
if self._stream is not None:
self._stream.close()
self._stream = None
@@ -0,0 +1,333 @@
#!/usr/bin/env python
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Standalone ``rt/smpl`` publisher for the PICO headset (no gear_sonic / torch).
Reads 24 body-joint poses from the XRoboToolkit SDK, runs pure-numpy SMPL forward
kinematics + canonicalization (``smpl_fk.py``), and publishes one canonical
``(24, 3)`` SMPL frame per tick over ZMQ on the ``rt/smpl`` topic — the exact
message ``lerobot.teleoperators.pico_headset.smpl_stream.SmplStream`` consumes.
This makes the LeRobot side self-contained: the only runtime dependency to drive
SONIC whole-body teleop from the headset is the ``xrobotoolkit_sdk`` Python package
(plus numpy/scipy/pyzmq), not the full ``gear_sonic`` stack.
Usage:
# Real headset (XRoboToolkit PC Service must be running and connected):
python -m lerobot.teleoperators.pico_headset.pico_publisher --fps 50 --port 5560
# No hardware — emit a synthetic waving motion to test the consumer end-to-end:
python -m lerobot.teleoperators.pico_headset.pico_publisher --fake
# Replay a canned SMPL clip to the robot through the same rt/smpl -> SONIC path:
python -m lerobot.teleoperators.pico_headset.pico_publisher \
--motion-file examples/unitree_g1/motions/walk_forward.npz
"""
from __future__ import annotations
import argparse
import contextlib
import json
import time
import numpy as np
import zmq
from lerobot.teleoperators.pico_headset.smpl_fk import (
SmplForwardKinematics,
ThreePointCalibrator,
canonicalize_smpl_joints,
compute_3point,
compute_3point_from_devices,
root_quats_from_aa,
)
SMPL_TOPIC = "rt/smpl"
DEFAULT_SMPL_PORT = 5560
def pack_message(
smpl_joints_local: np.ndarray,
frame_index: int,
stamp_ns: int,
root_quat: np.ndarray | None = None,
root_transl: np.ndarray | None = None,
vr3_pos: np.ndarray | None = None,
vr3_orn: np.ndarray | None = None,
loco_axes: np.ndarray | None = None,
loco_buttons: np.ndarray | None = None,
) -> bytes:
"""Build the rt/smpl JSON message (single frame, topic embedded in payload).
Carries the SMPL whole-body window (``smpl_joints_local`` + ``root_*``) and,
when available, the sparse 3-point VR targets (``vr3_pos`` (9,), ``vr3_orn`` (12,))
so a single stream can drive either SONIC ``encode_mode`` 1 or 2.
"""
data = {
"smpl_joints_local": np.asarray(smpl_joints_local, np.float32).reshape(-1).tolist(),
"frame_index": int(frame_index),
"stamp_ns": int(stamp_ns),
}
if root_quat is not None:
data["root_quat"] = np.asarray(root_quat, np.float32).reshape(-1).tolist()
if root_transl is not None:
data["root_transl"] = np.asarray(root_transl, np.float32).reshape(-1).tolist()
if vr3_pos is not None:
data["vr3_pos"] = np.asarray(vr3_pos, np.float32).reshape(-1).tolist()
if vr3_orn is not None:
data["vr3_orn"] = np.asarray(vr3_orn, np.float32).reshape(-1).tolist()
if loco_axes is not None:
data["loco_axes"] = np.asarray(loco_axes, np.float32).reshape(-1).tolist()
if loco_buttons is not None:
data["loco_buttons"] = np.asarray(loco_buttons, np.float32).reshape(-1).tolist()
return json.dumps({"topic": SMPL_TOPIC, "data": data}).encode("utf-8")
def _fake_body_poses(t: float) -> np.ndarray:
"""Synthetic (24, 7) body poses: identity rotations + a gently waving right arm."""
poses = np.zeros((24, 7), np.float32)
poses[:, 6] = 1.0 # unit quaternion (qw = 1), scalar-last
poses[:, 1] = 1.0 # ~1 m pelvis height (positions only matter for root_transl)
# Wave the right shoulder (SMPL body joint 17) about Z.
ang = 0.5 * np.sin(2.0 * np.pi * 0.5 * t)
poses[17, 3:7] = [0.0, 0.0, np.sin(ang / 2), np.cos(ang / 2)]
return poses
def _load_motion_clip(path: str) -> dict:
"""Load an SMPL ``.npz`` clip and canonicalize it for rt/smpl streaming.
Expects the same keys as ``motion_loader.SmplMotion``:
smpl_joints (T, 24, 3), pose_aa (T, 72) optional, transl (T, 3) optional.
Returns per-frame joints already in the encoder's root-removed convention,
plus optional per-frame root quat/translation.
"""
data = np.load(path)
joints = data["smpl_joints"].astype(np.float32)
if joints.ndim != 3 or joints.shape[1:] != (24, 3):
raise ValueError(f"Expected smpl_joints (T, 24, 3), got {joints.shape}")
pose_aa = data["pose_aa"].astype(np.float32) if "pose_aa" in data.files else None
root_quat = None
if pose_aa is not None:
joints = canonicalize_smpl_joints(joints, pose_aa[:, :3])
root_quat = root_quats_from_aa(pose_aa[:, :3])
transl = data["transl"].astype(np.float32) if "transl" in data.files else None
return {"joints": joints, "root_quat": root_quat, "transl": transl}
def main() -> None:
p = argparse.ArgumentParser(description=__doc__)
p.add_argument("--port", type=int, default=DEFAULT_SMPL_PORT, help="ZMQ PUB port for rt/smpl")
p.add_argument("--fps", type=float, default=50.0, help="Target publish rate (Hz)")
p.add_argument("--skeleton", type=str, default=None, help="Path to smpl_skeleton.npz")
p.add_argument(
"--headset-source",
choices=["body", "devices"],
default="body",
help=(
"Live headset keypoint source: 'body' uses full-body tracking "
"(get_body_joints_pose, needs PICO Motion Trackers) and drives both SMPL "
"(encode_mode 2) and 3-point; 'devices' uses head + 2 controllers only "
"(get_headset_pose + get_*_controller_pose, no trackers) and emits 3-point "
"(encode_mode 1) exclusively."
),
)
src = p.add_mutually_exclusive_group()
src.add_argument("--fake", action="store_true", help="Publish synthetic motion (no headset)")
src.add_argument("--motion-file", type=str, default=None, help="Replay an SMPL .npz clip over rt/smpl")
p.add_argument("--no-loop", action="store_true", help="Play a --motion-file once, then stop")
args = p.parse_args()
clip = _load_motion_clip(args.motion_file) if args.motion_file else None
# FK is only needed for live/synthetic (24,7) body poses; clips are pre-canonical.
fk = None
if clip is None:
fk = SmplForwardKinematics(args.skeleton) if args.skeleton else SmplForwardKinematics()
xrt = None
if clip is None and not args.fake:
try:
import xrobotoolkit_sdk as xrt # noqa: PLC0415
except ImportError as e:
raise SystemExit(
"xrobotoolkit_sdk not available. Install it, or run with --fake / --motion-file "
"to test the pipeline without a headset."
) from e
xrt.init()
print("[pico_publisher] XRoboToolkit initialized")
ctx = zmq.Context.instance()
sock = ctx.socket(zmq.PUB)
sock.bind(f"tcp://*:{args.port}")
src_desc = (
f"motion-file {args.motion_file}"
if clip
else ("fake" if args.fake else f"headset:{args.headset_source}")
)
print(
f"[pico_publisher] '{SMPL_TOPIC}' bound to tcp://*:{args.port} @ {args.fps:.0f} Hz "
f"[source: {src_desc}]"
)
if clip is not None:
print(f"[pico_publisher] clip frames={clip['joints'].shape[0]} loop={not args.no_loop}")
period = 1.0 / max(1.0, args.fps)
frame_index = 0
last_tracked = -1
t0 = time.time()
# 3-point operator calibration (device source only): map the operator's neutral
# rest pose onto the G1's neutral stance. Trigger a (re)capture with the A+B+X+Y
# controller combo, mirroring gear_sonic's ThreePointPose.calibrate_now.
# Device source: the "neck" is the headset (pitches when looking down), so keep the
# wrist targets in the yaw-local world frame rather than de-rotating by head tilt.
calibrator = (
ThreePointCalibrator(neck_relative_wrists=False)
if (xrt is not None and args.headset_source == "devices")
else None
)
calib_combo_last = False
if calibrator is not None:
print(
"[pico_publisher] 3-point calibration: stand in a neutral rest pose and press "
"A+B+X+Y on the controllers to (re)calibrate."
)
try:
while True:
loop_start = time.time()
vr3_pos = vr3_orn = None
loco_axes = loco_buttons = None
if clip is not None:
n = clip["joints"].shape[0]
if args.no_loop and frame_index >= n:
print("\n[pico_publisher] clip finished")
break
i = frame_index % n
joints = clip["joints"][i]
root_quat = None if clip["root_quat"] is None else clip["root_quat"][i]
root_transl = None if clip["transl"] is None else clip["transl"][i]
stamp_ns = time.time_ns()
elif not args.fake and args.headset_source == "devices":
# Controller-state 3-point path: head + 2 controllers only, no PICO
# Motion Trackers / body tracking. Emits encode_mode-1 targets only;
# the SMPL whole-body window is left as a zero placeholder.
head = np.asarray(xrt.get_headset_pose(), np.float32)
lc = np.asarray(xrt.get_left_controller_pose(), np.float32)
rc = np.asarray(xrt.get_right_controller_pose(), np.float32)
stamp_ns = int(xrt.get_time_stamp_ns())
if head.shape != (7,) or lc.shape != (7,) or rc.shape != (7,):
time.sleep(0.005)
continue
last_tracked = int(sum(np.linalg.norm(d[3:7]) > 1e-6 for d in (head, lc, rc)))
# Empty SMPL window: this source drives encode_mode 1 only, so the
# consumer must not mistake it for a (zero) whole-body reference.
joints = np.zeros((0, 3), np.float32)
root_quat = None
root_transl = None
vr3_pos, vr3_orn = compute_3point_from_devices(head, lc, rc)
# Edge-triggered (re)calibration on the A+B+X+Y combo.
combo_now = bool(
xrt.get_A_button() and xrt.get_B_button() and xrt.get_X_button() and xrt.get_Y_button()
)
if combo_now and not calib_combo_last:
calibrator.capture(vr3_pos, vr3_orn)
print("\n[pico_publisher] 3-point calibration captured (neutral pose).")
calib_combo_last = combo_now
vr3_pos, vr3_orn = calibrator.apply(vr3_pos, vr3_orn)
# Controller-stick locomotion (encode_mode 1, replicated): left/right
# sticks + A/B/X/Y. The all-four combo is reserved for calibration, so
# suppress the button pairs on that frame to avoid a spurious mode cycle.
la = np.asarray(xrt.get_left_axis(), np.float32).reshape(-1)
ra = np.asarray(xrt.get_right_axis(), np.float32).reshape(-1)
loco_axes = np.array([la[0], la[1], ra[0], ra[1]], np.float32)
btn = (
np.zeros(4, np.float32)
if combo_now
else np.array(
[
float(xrt.get_A_button()),
float(xrt.get_B_button()),
float(xrt.get_X_button()),
float(xrt.get_Y_button()),
],
np.float32,
)
)
loco_buttons = btn
else:
if args.fake:
body_poses = _fake_body_poses(loop_start - t0)
stamp_ns = time.time_ns()
else:
body_poses = np.asarray(xrt.get_body_joints_pose(), np.float32)
stamp_ns = int(xrt.get_time_stamp_ns())
if body_poses.shape != (24, 7):
time.sleep(0.005)
continue
# How many joints are actually being tracked (non-zero-norm quat).
# If this stays near 0, the headset isn't streaming body data (e.g.
# "Full body"/"Send" not enabled, trackers uncalibrated, or a test
# device) and the reference will be static regardless of your motion.
# In that case, use --headset-source devices for head+controllers only.
quat_norms = np.linalg.norm(body_poses[:, 3:7], axis=1)
last_tracked = int(np.count_nonzero(quat_norms > 1e-6))
out = fk.compute(body_poses)
joints = out["smpl_joints_local"]
root_quat = out["root_quat"]
root_transl = out["root_transl"]
# Also emit the sparse 3-point VR targets so the same stream can
# drive encode_mode 1 (3-point teleop) without a second producer.
vr3_pos, vr3_orn = compute_3point(body_poses)
sock.send(
pack_message(
joints,
frame_index,
stamp_ns,
root_quat=root_quat,
root_transl=root_transl,
vr3_pos=vr3_pos,
vr3_orn=vr3_orn,
loco_axes=loco_axes,
loco_buttons=loco_buttons,
)
)
frame_index += 1
if frame_index % int(max(1, args.fps)) == 0:
denom = 3 if (not args.fake and clip is None and args.headset_source == "devices") else 24
unit = "devices" if denom == 3 else "joints"
extra = f" | tracked {last_tracked}/{denom} {unit}" if last_tracked >= 0 else ""
if calibrator is not None:
extra += " | calibrated" if calibrator.is_calibrated else " | UNCALIBRATED"
print(f"[pico_publisher] sent {frame_index} frames{extra}", end="\r")
dt = time.time() - loop_start
if dt < period:
time.sleep(period - dt)
except KeyboardInterrupt:
print("\n[pico_publisher] stopping")
finally:
sock.close(0)
if xrt is not None:
with contextlib.suppress(Exception):
xrt.close()
if __name__ == "__main__":
main()
@@ -0,0 +1,80 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Single source of truth for the SONIC SMPL whole-body action protocol.
These constants define the wire format shared between the PICO teleoperator
(producer, ``pico_headset.py``), the offline motion->dataset converter
(``smpl_to_dataset.py``), the live stream (``smpl_stream.py``), and the Unitree G1
``SonicWholeBodyController`` (consumer). Keeping them here avoids the producer and
consumer silently drifting apart.
"""
# SMPL encoder window geometry (matches ``smpl_joints_10frame_step1``).
WINDOW = 10 # frames per encoder window
N_JOINTS = 24 # SMPL joints per frame
JOINT_DIM = 3 # xyz per joint
SMPL_OBS_DIM = WINDOW * N_JOINTS * JOINT_DIM # 720
# Flat action-dict keys carrying the reference through the standard lerobot action
# pipeline as scalar floats: ``smpl.0 .. smpl.719`` and ``root.0 .. root.3``.
SMPL_ACTION_PREFIX = "smpl."
ROOT_ACTION_PREFIX = "root."
ROOT_ACTION_DIM = 4 # per-frame SMPL root orientation (wxyz)
# Full per-frame action vector: 720 joint window + 4 root quaternion = 724.
ACTION_DIM = SMPL_OBS_DIM + ROOT_ACTION_DIM
# ── 3-point VR teleop protocol (SONIC encode_mode 1) ─────────────────────────
# An alternative, sparse upper-body reference: 3 root-relative keypoints
# (left wrist, right wrist, neck), each a position + orientation. The lower body /
# locomotion is driven by the planner (joystick/keyboard), not by these targets.
VR3_N_POINTS = 3 # left wrist, right wrist, neck
VR3_POS_DIM = VR3_N_POINTS * 3 # 9 (3 x xyz)
VR3_ORN_DIM = VR3_N_POINTS * 4 # 12 (3 x wxyz)
# Flat action-dict keys: ``vr3_pos.0 .. vr3_pos.8`` and ``vr3_orn.0 .. vr3_orn.11``.
VR3_POS_PREFIX = "vr3_pos."
VR3_ORN_PREFIX = "vr3_orn."
# ── Controller-stick locomotion (SONIC encode_mode 1, replicated exactly) ────
# In the original 3-point teleop the same tick that sends the VR targets also drives
# locomotion from the PICO controller sticks/buttons (left stick -> move, right stick
# -> yaw, A+B / X+Y -> cycle locomotion mode). We forward that raw controller state so
# the consumer's planner can steer walking/turning underneath the upper-body tracking.
LOCO_N_AXES = 4 # [left_x, left_y, right_x, right_y]
LOCO_N_BTN = 4 # [A, B, X, Y]
LOCO_AXES_PREFIX = "loco_axes."
LOCO_BTN_PREFIX = "loco_btn."
# ── Dense whole-body joint reference (SONIC encode_mode 0, OpenHLM-style) ─────
# A single 34-D whole-body command per tick, in the pi0.5 / OpenHLM action layout:
# [L-arm(7), L-grip(1), R-arm(7), R-grip(1), L-leg(6), R-leg(6), waist(3),
# root roll/pitch + yaw-rate(3)]
# The 29 joint targets (arms/legs/waist, grippers excluded) become the mode-0
# encoder joint reference and root roll/pitch become the anchor orientation.
#
# Fed as flat scalars ``wb.0.pos .. wb.33.pos``. The ``.pos`` suffix is required so
# these behave like ordinary joint-position action features: ``lerobot-rollout``
# only routes ``*.pos`` keys from ``robot.action_features`` into the policy<->robot
# action mapping, letting a 34-D VLA (OpenHLM / pi0.5) drive the robot directly.
WB_ACTION_PREFIX = "wb."
WB_ACTION_DIM = 34
def wb_action_key(i: int) -> str:
"""Action-dict key for the ``i``-th whole-body command scalar (``wb.{i}.pos``)."""
return f"{WB_ACTION_PREFIX}{i}.pos"
@@ -0,0 +1,571 @@
#!/usr/bin/env python
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Standalone SMPL forward kinematics + canonicalization in pure numpy/scipy.
This mirrors the ``rt/smpl`` producer path in ``gear_sonic`` (``compute_from_body_poses``
-> ``process_smpl_joints`` -> ``compute_human_joints``) without depending on torch,
the SMPL mesh model, or ``gear_sonic``. It only needs a small fixed skeleton table
(SMPL-X rest-pose joints + kinematic tree), hardcoded below as ``_SKELETON_J`` /
``_SKELETON_PARENTS`` so no external asset download is required.
Given the 24 body-joint poses reported by the XRoboToolkit headset SDK
(``xrt.get_body_joints_pose()`` -> (24, 7) of ``[x, y, z, qx, qy, qz, qw]``), it
produces the root-orientation-removed 24x3 SMPL joints the SONIC encoder expects,
plus the root orientation quaternion and pelvis translation.
Quaternions are scalar-first (w, x, y, z) unless noted.
"""
from dataclasses import dataclass, field
from pathlib import Path
import numpy as np
from scipy.spatial.transform import Rotation as R # noqa: N817
from .smpl_constants import VR3_N_POINTS, VR3_ORN_DIM, VR3_POS_DIM
# 24-joint parent tree used by the headset body-pose stream (SMPL-X body subset).
# Matches PoseStreamer.parent_indices in gear_sonic's pico_manager_thread_server.py.
BODY24_PARENTS = np.array(
[-1, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 9, 9, 12, 13, 14, 16, 17, 18, 19, 20, 22],
dtype=np.int64,
)
# FK output joints: first 22 SMPL body joints + two thumb tips (SMPL-X indices 39, 54).
OUTPUT_JOINT_INDEX = np.concatenate([np.arange(22), np.array([39, 54])])
# SMPL-X rest-pose skeleton (55 joints), vendored inline to avoid an external
# asset. ``_SKELETON_PARENTS[i]`` is joint ``i``'s parent (-1 = root); ``_SKELETON_J``
# holds the (55, 3) rest-pose joint positions.
_SKELETON_PARENTS = np.array(
[
-1,
0,
0,
0,
1,
2,
3,
4,
5,
6,
7,
8,
9,
9,
9,
12,
13,
14,
16,
17,
18,
19,
15,
15,
15,
20,
25,
26,
20,
28,
29,
20,
31,
32,
20,
34,
35,
20,
37,
38,
21,
40,
41,
21,
43,
44,
21,
46,
47,
21,
49,
50,
21,
52,
53,
],
dtype=np.int64,
)
_SKELETON_J = np.array(
[
[0.0031232605688273907, -0.3514074683189392, 0.012036550790071487],
[0.06131265312433243, -0.4441709518432617, -0.013964635320007801],
[-0.06014421582221985, -0.4553154706954956, -0.009213820099830627],
[0.00036056205863133073, -0.2415168583393097, -0.015581080690026283],
[0.11600811034440994, -0.8229243755340576, -0.02336069941520691],
[-0.10435417294502258, -0.8176955580711365, -0.026037702336907387],
[0.009808260947465897, -0.10966360569000244, -0.02152106538414955],
[0.07255466282367706, -1.2259838581085205, -0.05523664504289627],
[-0.08893736451864243, -1.2284233570098877, -0.046229973435401917],
[-0.0015221529174596071, -0.057428449392318726, 0.006925832014530897],
[0.11981196701526642, -1.283981204032898, 0.06297968327999115],
[-0.12774977087974548, -1.2867517471313477, 0.07281902432441711],
[-0.01368661131709814, 0.10773860663175583, -0.024689510464668274],
[0.04484200477600098, 0.027515273541212082, -0.0002946509048342705],
[-0.04921707883477211, 0.026910223066806793, -0.006474069785326719],
[0.011096873320639133, 0.2681904137134552, -0.003952245227992535],
[0.16408103704452515, 0.08524329960346222, -0.015755590051412582],
[-0.15179482102394104, 0.08043467253446579, -0.019142597913742065],
[0.4182038903236389, 0.01309278141707182, -0.058214444667100906],
[-0.4229443669319153, 0.04394219070672989, -0.04560968279838562],
[0.6701906323432922, 0.03631401062011719, -0.06068652495741844],
[-0.6722118258476257, 0.03940964490175247, -0.06093486770987511],
[-0.004667762666940689, 0.2676706910133362, -0.009591402485966682],
[0.03159928321838379, 0.31083211302757263, 0.062195174396038055],
[-0.031599875539541245, 0.3108319342136383, 0.0621943436563015],
[0.7720924615859985, 0.02762586995959282, -0.04133538901805878],
[0.8040408492088318, 0.02984413132071495, -0.03832494467496872],
[0.8265857696533203, 0.027494050562381744, -0.03827037289738655],
[0.7795881628990173, 0.029986342415213585, -0.06466733664274216],
[0.8101993799209595, 0.030793743208050728, -0.0686899945139885],
[0.8337147831916809, 0.028784994035959244, -0.07280422002077103],
[0.7542374730110168, 0.02177468128502369, -0.10443613678216934],
[0.7697082161903381, 0.020643100142478943, -0.11643557250499725],
[0.7852417230606079, 0.018978532403707504, -0.12765252590179443],
[0.767634928226471, 0.02704637683928013, -0.08803117275238037],
[0.7956817150115967, 0.028531836345791817, -0.09329714626073837],
[0.8185034990310669, 0.02707355096936226, -0.1003914326429367],
[0.7108263969421387, 0.01833728514611721, -0.03507564589381218],
[0.7278420925140381, 0.01931309886276722, -0.010097505524754524],
[0.7483652234077454, 0.01415354385972023, 0.005425570998340845],
[-0.7720924019813538, 0.027626780793070793, -0.041334930807352066],
[-0.8040405511856079, 0.029844673350453377, -0.03832409530878067],
[-0.8265854716300964, 0.027495287358760834, -0.03826868534088135],
[-0.7795882225036621, 0.029987698420882225, -0.0646686926484108],
[-0.8101993799209595, 0.030795171856880188, -0.06869153678417206],
[-0.8337149024009705, 0.02878585271537304, -0.07280556112527847],
[-0.7542385458946228, 0.021775206550955772, -0.10443780571222305],
[-0.7697089910507202, 0.02064313367009163, -0.11643654853105545],
[-0.7852423787117004, 0.01897839829325676, -0.1276528388261795],
[-0.7676352858543396, 0.02704770304262638, -0.08803359419107437],
[-0.7956817150115967, 0.028532907366752625, -0.093299500644207],
[-0.818503737449646, 0.027074117213487625, -0.10039224475622177],
[-0.7108249664306641, 0.018335221335291862, -0.035073522478342056],
[-0.7278403043746948, 0.019311318174004555, -0.01009594276547432],
[-0.7483659386634827, 0.014154116623103619, 0.005425604991614819],
],
dtype=np.float32,
)
# ── fixed frame corrections (shared by FK + canonicalization) ────────────────
# Both mirror the SONIC deploy transform. ``_YTOZ_UP`` maps SMPL's Y-up world to the
# robot's Z-up (a 90 deg rotation about X); ``_SMPL_BASE`` is SMPL's rest-pose base
# orientation, conjugated out during canonicalization.
_YTOZ_UP = R.from_euler("x", 90, degrees=True)
_SMPL_BASE = R.from_quat([0.5, 0.5, 0.5, 0.5]) # scalar-last; symmetric so wxyz==xyzw
# ── forward kinematics ───────────────────────────────────────────────────────
def canonicalize_smpl_joints(smpl_joints: np.ndarray, root_aa: np.ndarray) -> np.ndarray:
"""Remove per-frame root orientation -> SONIC ``smpl_joints_local`` format.
Mirrors the deploy transform (and ``motion_loader.canonicalize_smpl_joints``):
reference clips store world-frame joints, but the encoder wants each frame's
joints with the body root orientation removed.
Args:
smpl_joints: (T, 24, 3) world-frame (z-up) SMPL joint positions.
root_aa: (T, 3) SMPL global-orient axis-angle (y-up convention).
Returns:
(T, 24, 3) per-frame root-orientation-removed joints.
"""
root = _YTOZ_UP * R.from_rotvec(root_aa)
inv = _SMPL_BASE * root.inv()
return np.einsum("tij,tkj->tki", inv.as_matrix(), smpl_joints).astype(np.float32)
def root_quats_from_aa(root_aa: np.ndarray) -> np.ndarray:
"""Per-frame root orientation as (T, 4) wxyz, matching the live ``root_quat``.
Same convention as the headset stream: ytoz-up then base-rotation removed.
"""
root = (_YTOZ_UP * R.from_rotvec(root_aa)) * _SMPL_BASE.inv()
return root.as_quat(scalar_first=True).astype(np.float32) # wxyz
# ── 3-point VR teleop keypoints (SONIC encode_mode 1) ────────────────────────
# SMPL body-joint indices for the 3 tracked keypoints, plus the root/pelvis (0)
# used as the reference frame. Mirrors gear_sonic ``_process_3pt_pose``: neck
# (joint 12) is used rather than head (15) — it is more rigidly coupled to the
# torso and less noisy than the free-looking head.
_VR3_JOINTS = (22, 23, 12) # left wrist, right wrist, neck
# Per-keypoint rotation offsets aligning each SMPL joint frame to the robot
# convention (root, left wrist, right wrist, neck), ported verbatim from
# gear_sonic ``pico_manager_thread_server.OFFSETS`` (extrinsic xyz euler, degrees).
_VR3_OFFSETS = [
R.from_euler("xyz", [0, 0, -90], degrees=True), # root
R.from_euler("xyz", [90, 0, 0], degrees=True), # left wrist
R.from_euler("xyz", [-90, 0, 180], degrees=True), # right wrist
R.from_euler("xyz", [0, 0, -90], degrees=True), # neck
]
# Unity (X-right, Y-up, Z-forward, left-handed) -> robot (X-forward, Y-left,
# Z-up, right-handed) axis remap: Unity [x, y, z] -> robot [-x, z, y].
_UNITY_TO_ROBOT = np.array([[-1.0, 0.0, 0.0], [0.0, 0.0, 1.0], [0.0, 1.0, 0.0]])
def _safe_quat(quats: np.ndarray) -> np.ndarray:
"""Replace zero-norm quaternions with the scalar-last identity.
The headset reports ``[0, 0, 0, 0]`` for joints it isn't currently tracking;
``scipy.Rotation.from_quat`` rejects zero-norm quaternions, so we substitute the
identity ``[0, 0, 0, 1]`` (no rotation) for those rows to keep FK robust.
"""
quats = np.asarray(quats, np.float64).copy()
bad = np.linalg.norm(quats, axis=-1) < 1e-8
quats[bad] = (0.0, 0.0, 0.0, 1.0) # scalar-last identity
return quats
def compute_3point(body_poses_np: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
"""Extract the SONIC 3-point VR targets from headset body poses.
Mirrors gear_sonic ``_process_3pt_pose``: transforms the tracked joints from the
Unity frame to the robot frame, applies the per-keypoint rotation offsets, then
expresses the left-wrist / right-wrist / neck keypoints relative to the root
(pelvis) frame. This is the ``encode_mode == 1`` counterpart to :func:`compute`.
Note: physical wrist/neck position offsets and the operator calibration done in
gear_sonic's ``ThreePointPose.apply_calibration`` are not applied here; the raw
tracked joint poses are used.
Args:
body_poses_np: (24, 7) rows of ``[x, y, z, qx, qy, qz, qw]`` (scalar-last),
in the Unity frame, as returned by ``xrt.get_body_joints_pose()``.
Returns:
(pos, orn):
- pos: (9,) float32, root-relative ``[x, y, z]`` for [l-wrist, r-wrist, neck]
- orn: (12,) float32, root-relative ``[w, x, y, z]`` for the same order
"""
body = np.asarray(body_poses_np, np.float64)
q = _UNITY_TO_ROBOT
# Root (index 0) + the 3 tracked keypoints, each transformed to the robot frame
# and rotation-offset-corrected.
positions = np.zeros((4, 3), np.float64)
rotations: list[R] = []
quats = _safe_quat(body[:, 3:7])
for out_i, j in enumerate((0, *_VR3_JOINTS)):
positions[out_i] = q @ body[j, :3]
rot = R.from_quat(quats[j]).as_matrix() # scalar-last input
rotations.append(R.from_matrix(q @ rot @ q.T) * _VR3_OFFSETS[out_i])
root_inv = rotations[0].inv()
root_pos = positions[0]
pos = np.zeros(VR3_POS_DIM, np.float32)
orn = np.zeros(VR3_ORN_DIM, np.float32)
for k in range(VR3_N_POINTS):
pos[k * 3 : k * 3 + 3] = root_inv.apply(positions[k + 1] - root_pos)
orn[k * 4 : k * 4 + 4] = (root_inv * rotations[k + 1]).as_quat(scalar_first=True) # wxyz
return pos, orn
# ── 3-point VR teleop from raw device poses (no body trackers) ───────────────
# PICO Y-up (X-right, Y-up, Z-back) -> robot Z-up world. Ported verbatim from
# gear_sonic's controller path (``decoupled_wbc`` ``PicoStreamer.R_HEADSET_TO_WORLD``).
_HEADSET_TO_WORLD = np.array([[0.0, 0.0, -1.0], [-1.0, 0.0, 0.0], [0.0, 1.0, 0.0]])
def _device_pose_to_world(pose: np.ndarray) -> tuple[np.ndarray, R]:
"""Convert a raw (7,) device pose ``[x, y, z, qx, qy, qz, qw]`` (PICO Y-up frame)
to a Z-up world ``(position, Rotation)`` pair.
Handles the all-zero quaternion the SDK emits when a device is momentarily
untracked by substituting the identity, matching ``PicoStreamer._process_xr_pose``.
"""
pose = np.asarray(pose, np.float64)
xyz = _HEADSET_TO_WORLD @ pose[:3]
quat = pose[3:7] # scalar-last
if np.linalg.norm(quat) < 1e-8:
quat = np.array([0.0, 0.0, 0.0, 1.0])
rot = _HEADSET_TO_WORLD @ R.from_quat(quat).as_matrix() @ _HEADSET_TO_WORLD.T
return xyz, R.from_matrix(rot)
def compute_3point_from_devices(
head_pose: np.ndarray,
left_pose: np.ndarray,
right_pose: np.ndarray,
) -> tuple[np.ndarray, np.ndarray]:
"""Build the SONIC 3-point VR targets from raw head + controller poses.
This is the controller-state path (no PICO Motion Trackers / body tracking
required): the 3 keypoints are the left controller, right controller, and the
headset, each expressed relative to the **headset yaw frame** — mirroring
gear_sonic's ``decoupled_wbc`` ``PicoStreamer._process_xr_pose`` (Y-up -> Z-up,
then de-headed by the headset yaw). The headset stands in for the "neck" point,
so its root-relative position is ~0 and its orientation carries pitch/roll.
Args:
head_pose, left_pose, right_pose: (7,) ``[x, y, z, qx, qy, qz, qw]`` device
poses (scalar-last) from ``xrt.get_headset_pose()`` /
``xrt.get_left_controller_pose()`` / ``xrt.get_right_controller_pose()``.
Returns:
(pos, orn):
- pos: (9,) float32, headset-yaw-relative ``[x, y, z]`` for [l-wrist, r-wrist, head]
- orn: (12,) float32, headset-yaw-relative ``[w, x, y, z]`` for the same order
"""
head_pos, head_rot = _device_pose_to_world(head_pose)
left_pos, left_rot = _device_pose_to_world(left_pose)
right_pos, right_rot = _device_pose_to_world(right_pose)
# De-head: cancel the headset yaw so targets are expressed in a heading-local frame.
head_yaw = head_rot.as_euler("xyz")[2]
inv_yaw = R.from_euler("z", -head_yaw)
points = ((left_pos, left_rot), (right_pos, right_rot), (head_pos, head_rot))
pos = np.zeros(VR3_POS_DIM, np.float32)
orn = np.zeros(VR3_ORN_DIM, np.float32)
for k, (p_pos, p_rot) in enumerate(points):
pos[k * 3 : k * 3 + 3] = inv_yaw.apply(p_pos - head_pos)
orn[k * 4 : k * 4 + 4] = (inv_yaw * p_rot).as_quat(scalar_first=True) # wxyz
return pos, orn
# ── operator calibration for the 3-point targets ────────────────────────────
# G1 neutral key-frame targets, pelvis-relative [x, y, z] in metres, from MuJoCo FK on
# g1_29dof at the robot's *standing* configuration (``default_angles`` — the pose the
# robot actually holds at calibration time), with gear_sonic's local key-frame offsets
# (``G1_KEY_FRAME_OFFSETS``: wrists +0.18x ∓0.025y, torso +0.35z). These are the poses
# the operator's rest pose is mapped onto so the handoff starts at the robot's neutral
# stance. This is the fixed-``default_angles`` stand-in for gear_sonic's
# ``get_g1_key_frame_poses(q=measured_q)`` (see :meth:`ThreePointCalibrator.capture`):
# since the robot stands at ``default_angles`` after the startup ramp, its measured q
# equals this configuration, so these constants are the measured-q targets for the
# nominal case. (Per-frame *live* measured q would need a reverse controller->publisher
# feedback channel; not wired.)
_G1_NEUTRAL_WRIST_POS = np.array([[0.2232, 0.2177, -0.1555], [0.2232, -0.2177, -0.1555]], np.float64)
# Wrist neutral orientations (scalar-first w, x, y, z) at ``default_angles`` — NOT
# identity: the wrists are rolled/pitched in the standing pose. Matches gear_sonic
# using ``g1_lwrist_rot`` / ``g1_rwrist_rot`` from FK (not identity) as the rotation
# calibration reference.
_G1_NEUTRAL_WRIST_ROT = [
R.from_quat([0.9168, 0.0897, 0.3864, 0.0463], scalar_first=True), # left
R.from_quat([0.9168, -0.0897, 0.3864, -0.0463], scalar_first=True), # right
]
# Neck reconstruction chain (mirrors ThreePointPose._apply_calibration): torso link
# +0.05 z, then +0.35 along the neck's local Z.
_NECK_TORSO_OFFSET_Z = 0.05
_NECK_LINK_LENGTH = 0.35
@dataclass
class ThreePointCalibrator:
"""Aligns raw 3-point VR targets to the G1's neutral stance.
Ports gear_sonic ``ThreePointPose._capture_calibration`` / ``_apply_calibration``:
on :meth:`capture` (operator holding a neutral rest pose) it records (a) the
inverse of the head/neck orientation, used to de-tilt all points to upright, and
(b) per-wrist position + orientation offsets that map the corrected rest pose onto
the fixed G1 neutral wrist targets. :meth:`apply` then transforms every subsequent
frame by those offsets, and reconstructs the head/neck position from the calibrated
neck orientation via the torso->neck kinematic chain.
All quaternions are scalar-first (w, x, y, z), matching :func:`compute_3point`.
``neck_relative_wrists`` selects how the wrist targets are framed:
- ``True`` (body-source, :func:`compute_3point`): wrists are pelvis-relative and
the neck sits roughly upright over the pelvis, so de-rotating them by ``neck_inv``
correctly expresses them in the neck/torso frame (gear_sonic's behaviour).
- ``False`` (device-source, :func:`compute_3point_from_devices`): the "neck" is the
**headset**, which pitches down when the operator looks at their hands; the wrists
are already yaw-stabilised, so applying the head pitch/roll would rotate "up" hand
motion into "forward". The wrist frame is left in the yaw-local world frame (only
the neck point itself is still de-tilted).
"""
neck_relative_wrists: bool = True
_neck_quat_inv: R | None = field(default=None, init=False)
_wrist_pos_offset: np.ndarray | None = field(default=None, init=False)
_wrist_rot_offset: list[R] = field(default_factory=list, init=False)
@property
def is_calibrated(self) -> bool:
return self._neck_quat_inv is not None and self._wrist_pos_offset is not None
def reset(self) -> None:
self._neck_quat_inv = None
self._wrist_pos_offset = None
self._wrist_rot_offset = []
def recalibrate_wrists(self) -> None:
"""Clear only the wrist offsets, preserving the neck calibration.
Mirrors gear_sonic ``ThreePointPose.reset_with_measured_q``: the next
:meth:`capture` recomputes the wrist offsets (against the G1 neutral targets)
while keeping the already-captured neck orientation, so re-aligning the arms
doesn't force the operator to re-level their head.
"""
self._wrist_pos_offset = None
self._wrist_rot_offset = []
def capture(self, pos: np.ndarray, orn: np.ndarray) -> None:
"""Capture calibration offsets from a neutral-pose frame.
Neck calibration is captured once and then preserved across subsequent
captures (matching gear_sonic's ``if self._calibration_neck_quat_inv is
None``); call :meth:`reset` to clear it or :meth:`recalibrate_wrists` to
re-align only the arms.
Args:
pos: (9,) root-relative ``[x, y, z]`` for [l-wrist, r-wrist, head].
orn: (12,) root-relative ``[w, x, y, z]`` for the same order.
"""
pos = np.asarray(pos, np.float64).reshape(3, 3)
orn = np.asarray(orn, np.float64).reshape(3, 4)
if self._neck_quat_inv is None:
self._neck_quat_inv = R.from_quat(orn[2], scalar_first=True).inv()
# Wrists use the neck frame only in body-source mode; device-source keeps them
# in the already yaw-stabilised world frame (see class docstring).
wrist_inv = self._neck_quat_inv if self.neck_relative_wrists else R.identity()
self._wrist_pos_offset = np.zeros((2, 3), np.float64)
self._wrist_rot_offset = []
for k in range(2):
corrected_pos = wrist_inv.apply(pos[k])
corrected_rot = wrist_inv * R.from_quat(orn[k], scalar_first=True)
self._wrist_pos_offset[k] = corrected_pos - _G1_NEUTRAL_WRIST_POS[k]
# rot_offset maps the corrected rest orientation onto the G1 neutral wrist
# orientation: calibrated = rot_offset * (neck_inv * current).
self._wrist_rot_offset.append(_G1_NEUTRAL_WRIST_ROT[k] * corrected_rot.inv())
def apply(self, pos: np.ndarray, orn: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
"""Apply the stored calibration; returns calibrated ``(pos (9,), orn (12,))``.
A no-op (returns the inputs unchanged) until :meth:`capture` has been called.
"""
if self._neck_quat_inv is None or self._wrist_pos_offset is None:
return (
np.asarray(pos, np.float32).reshape(-1),
np.asarray(orn, np.float32).reshape(-1),
)
pos = np.asarray(pos, np.float64).reshape(3, 3)
orn = np.asarray(orn, np.float64).reshape(3, 4)
neck_inv = self._neck_quat_inv
wrist_inv = neck_inv if self.neck_relative_wrists else R.identity()
out_pos = np.zeros((3, 3), np.float64)
out_orn = np.zeros((3, 4), np.float64)
for k in range(2): # wrists
out_pos[k] = wrist_inv.apply(pos[k]) - self._wrist_pos_offset[k]
corrected_rot = wrist_inv * R.from_quat(orn[k], scalar_first=True)
out_orn[k] = (self._wrist_rot_offset[k] * corrected_rot).as_quat(scalar_first=True)
# Head/neck: orientation de-tilted, position from the torso->neck chain.
neck_rot = neck_inv * R.from_quat(orn[2], scalar_first=True)
out_orn[2] = neck_rot.as_quat(scalar_first=True)
neck_z = neck_rot.apply([0.0, 0.0, 1.0])
out_pos[2] = np.array([0.0, 0.0, _NECK_TORSO_OFFSET_Z]) + _NECK_LINK_LENGTH * neck_z
return out_pos.reshape(-1).astype(np.float32), out_orn.reshape(-1).astype(np.float32)
class SmplForwardKinematics:
"""Rest-skeleton SMPL forward kinematics (no mesh, no torch)."""
def __init__(self, skeleton_path: str | Path | None = None):
if skeleton_path is not None:
data = np.load(skeleton_path)
self.J = data["J"].astype(np.float64) # (55, 3) rest joint positions
self.parents = data["parents"].astype(np.int64) # (55,) kinematic tree
else:
self.J = _SKELETON_J.astype(np.float64)
self.parents = _SKELETON_PARENTS.copy()
self.n_joints = self.J.shape[0]
def _fk(self, full_pose_aa: np.ndarray) -> np.ndarray:
"""full_pose_aa: (n_joints, 3) axis-angle (joint 0 = global). Returns (24, 3)."""
rot = R.from_rotvec(full_pose_aa).as_matrix() # (n, 3, 3)
rel = self.J.copy()
rel[1:] -= self.J[self.parents[1:]]
transforms = np.zeros((self.n_joints, 4, 4), np.float64)
transforms[:, :3, :3] = rot
transforms[:, :3, 3] = rel
transforms[:, 3, 3] = 1.0
chain = [transforms[0]]
for i in range(1, self.n_joints):
chain.append(chain[self.parents[i]] @ transforms[i])
joints = np.stack(chain)[:, :3, 3]
return joints[OUTPUT_JOINT_INDEX]
def compute(self, body_poses_np: np.ndarray) -> dict:
"""Convert (24, 7) headset body poses to canonical SMPL joints.
Args:
body_poses_np: (24, 7) rows of [x, y, z, qx, qy, qz, qw] (scalar-last).
Returns:
dict with:
- smpl_joints_local: (24, 3) root-orientation-removed joints
- root_quat: (4,) root/torso orientation (w, x, y, z)
- root_transl: (3,) pelvis translation
"""
body_poses_np = np.asarray(body_poses_np, np.float64)
positions = body_poses_np[:, :3]
# Global joint rotations from the headset (scalar-last), with the SMPL
# +180 deg-about-Y frame fix, converted to per-joint local axis-angle.
global_rots = R.from_quat(_safe_quat(body_poses_np[:, 3:7])) * R.from_euler("y", 180, degrees=True)
gm = global_rots.as_matrix() # (24, 3, 3)
local_aa = np.zeros((24, 3), np.float64)
for i in range(24):
p = BODY24_PARENTS[i]
m = gm[i] if p == -1 else gm[p].T @ gm[i]
local_aa[i] = R.from_matrix(m).as_rotvec()
global_orient = local_aa[0]
body_pose = local_aa[1:].reshape(-1)[:63] # 21 body joints
# Root: Y-up -> Z-up, then run FK with the transformed root.
root = _YTOZ_UP * R.from_rotvec(global_orient)
global_orient_new = root.as_rotvec()
full_pose = np.concatenate([global_orient_new, body_pose, np.zeros(3 * self.n_joints - 66)]).reshape(
self.n_joints, 3
)
joints = self._fk(full_pose) # (24, 3)
# Canonicalize: strip SMPL base rot and the root orientation.
root = root * _SMPL_BASE.inv()
smpl_joints_local = root.inv().apply(joints)
return {
"smpl_joints_local": smpl_joints_local.astype(np.float32),
"root_quat": root.as_quat(scalar_first=True).astype(np.float32), # wxyz
"root_transl": positions[0].astype(np.float32),
}
@@ -0,0 +1,277 @@
#!/usr/bin/env python
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Live SMPL stream as a SONIC reference motion (drop-in for ``SmplMotion``).
Instead of reading an ``.npz`` clip, this pulls per-frame SMPL joints live off the
``rt/smpl`` ZMQ channel published by the GEAR PICO teleop
(``gear_sonic/scripts/pico_manager_thread_server.py``). Each message carries one
frame of **canonical** (root-orientation-removed) SMPL local joints ``(24, 3)`` --
the exact per-frame format ``SmplMotion`` emits -- so this class exposes the same
``step() -> (720,)`` window interface and can be handed to ``sonic.py`` (or the
``pico_headset`` teleoperator) wherever a reference motion is expected
(``encode_mode == 2``).
Transport mirrors the Unitree SDK socket bridge (``unitree_sdk2_socket.py``): a ZMQ
``SUB`` socket with ``CONFLATE`` (keep only the latest frame) subscribed to the
``rt/smpl`` topic, JSON payloads.
"""
from __future__ import annotations
import contextlib
import json
import logging
import time
from collections import deque
import numpy as np
import zmq
from .smpl_constants import (
JOINT_DIM,
LOCO_N_AXES,
LOCO_N_BTN,
N_JOINTS,
SMPL_OBS_DIM,
VR3_ORN_DIM,
VR3_POS_DIM,
WINDOW,
)
logger = logging.getLogger(__name__)
SMPL_TOPIC = "rt/smpl"
DEFAULT_SMPL_HOST = "127.0.0.1"
DEFAULT_SMPL_PORT = 5560
class SmplStream:
"""Live ``rt/smpl`` consumer with the ``SmplMotion`` interface.
Args:
host: publisher host (the laptop running pico_manager_thread_server.py).
port: publisher port for the ``rt/smpl`` channel.
fps: nominal source rate, only used for status/reporting.
stale_after_s: log a warning if no fresh frame arrives within this window.
loop: accepted for API parity with ``SmplMotion`` (ignored; a stream never ends).
"""
def __init__(
self,
host: str = DEFAULT_SMPL_HOST,
port: int = DEFAULT_SMPL_PORT,
fps: float = 50.0,
stale_after_s: float = 0.5,
loop: bool = True,
):
self.host = host
self.port = port
self.fps = float(fps)
self.loop = loop
self.stale_after_s = stale_after_s
self._ctx = zmq.Context.instance()
self._sock = self._ctx.socket(zmq.SUB)
self._sock.setsockopt(zmq.CONFLATE, 1) # keep only the most recent frame
self._sock.connect(f"tcp://{host}:{port}")
# Single-frame JSON messages (topic embedded in payload); CONFLATE does not
# support multipart, so subscribe to everything on this dedicated port.
self._sock.setsockopt_string(zmq.SUBSCRIBE, "")
self._poller = zmq.Poller()
self._poller.register(self._sock, zmq.POLLIN)
# Rolling window, oldest -> newest (matches SmplMotion.window layout).
self._buf: deque[np.ndarray] = deque(maxlen=WINDOW)
self._last_frame = np.zeros((N_JOINTS, JOINT_DIM), np.float32)
# Latest root/torso pose (updated every received frame).
self.root_quat = np.array([1.0, 0.0, 0.0, 0.0], np.float32) # (w, x, y, z)
self.root_transl = np.zeros(3, np.float32)
# Latest sparse 3-point VR targets (encode_mode 1), if the producer sends them.
self.vr3_pos = np.zeros(VR3_POS_DIM, np.float32)
self.vr3_orn = np.tile([1.0, 0.0, 0.0, 0.0], VR3_ORN_DIM // 4).astype(np.float32)
self._got_vr3 = False
self._last_vr3_t = 0.0
# Latest controller-stick locomotion (encode_mode 1): [lx, ly, rx, ry] + [A,B,X,Y].
self.loco_axes = np.zeros(LOCO_N_AXES, np.float32)
self.loco_buttons = np.zeros(LOCO_N_BTN, np.float32)
self._got_loco = False
self._last_loco_t = 0.0
self._last_index = -1
self._last_recv_t = 0.0
self._warned_stale = False
self._got_first = False
# -- SmplMotion-compatible attributes ------------------------------------
@property
def num_frames(self) -> int:
"""Streams are unbounded; report 0 (kept for API parity)."""
return 0
@property
def done(self) -> bool:
"""A live stream never finishes."""
return False
@property
def has_data(self) -> bool:
"""True once at least one real frame has been received."""
return self._got_first
@property
def has_vr3(self) -> bool:
"""True once the producer has sent at least one 3-point VR frame."""
return self._got_vr3
@property
def has_fresh_vr3(self) -> bool:
"""True when a 3-point VR frame arrived within ``stale_after_s``.
Unlike :attr:`has_data`, this is independent of the SMPL window, so the
controller-state source (head + controllers only, empty SMPL) still drives
``encode_mode 1`` without a whole-body reference.
"""
if not self._got_vr3:
return False
if not self.stale_after_s:
return True
return (time.time() - self._last_vr3_t) <= self.stale_after_s
@property
def has_fresh_loco(self) -> bool:
"""True when controller-stick locomotion arrived within ``stale_after_s``."""
if not self._got_loco:
return False
if not self.stale_after_s:
return True
return (time.time() - self._last_loco_t) <= self.stale_after_s
@property
def seconds_since_last(self) -> float:
"""Wall-clock seconds since the last real frame (inf before the first)."""
if not self._got_first:
return float("inf")
return time.time() - self._last_recv_t
@property
def is_stale(self) -> bool:
"""True when the stream has gone silent past ``stale_after_s``.
Consumers use this to stop feeding a frozen pose and let the controller
fall back to a safe standing/locomotion mode.
"""
if not self._got_first or not self.stale_after_s:
return False
return self.seconds_since_last > self.stale_after_s
def reset(self):
self._buf.clear()
self._got_first = False
self._got_vr3 = False
self._last_vr3_t = 0.0
self._got_loco = False
self._last_loco_t = 0.0
# -- core ----------------------------------------------------------------
def _drain_latest(self) -> np.ndarray | None:
"""Return the newest available (24, 3) frame, or None if nothing new.
CONFLATE already keeps only the last message, but we poll non-blocking so
the 50 Hz control loop never stalls waiting on the headset.
"""
frame = None
while dict(self._poller.poll(0)).get(self._sock) == zmq.POLLIN:
payload = self._sock.recv()
data = json.loads(payload.decode("utf-8")).get("data", {})
# Sparse 3-point VR targets (encode_mode 1). Parsed independently of the
# SMPL window so the controller-state source (head + controllers, empty
# SMPL) is still handled.
vp = data.get("vr3_pos")
vo = data.get("vr3_orn")
if vp is not None and vo is not None and len(vp) == VR3_POS_DIM and len(vo) == VR3_ORN_DIM:
self.vr3_pos = np.asarray(vp, np.float32)
self.vr3_orn = np.asarray(vo, np.float32)
self._got_vr3 = True
self._last_vr3_t = time.time()
# Controller-stick locomotion (encode_mode 1), also independent of SMPL.
la = data.get("loco_axes")
lb = data.get("loco_buttons")
if la is not None and lb is not None and len(la) == LOCO_N_AXES and len(lb) == LOCO_N_BTN:
self.loco_axes = np.asarray(la, np.float32)
self.loco_buttons = np.asarray(lb, np.float32)
self._got_loco = True
self._last_loco_t = time.time()
# SMPL whole-body window (encode_mode 2), optional on this stream.
joints = np.asarray(data.get("smpl_joints_local", []), np.float32)
if joints.size != N_JOINTS * JOINT_DIM:
continue
frame = joints.reshape(N_JOINTS, JOINT_DIM)
self._last_index = int(data.get("frame_index", self._last_index + 1))
rq = data.get("root_quat")
if rq is not None and len(rq) == 4:
self.root_quat = np.asarray(rq, np.float32)
rt = data.get("root_transl")
if rt is not None and len(rt) == 3:
self.root_transl = np.asarray(rt, np.float32)
return frame
def step(self) -> np.ndarray:
"""Advance one control tick, returning the current 720-vec window.
If no new headset frame arrived this tick we hold the last one, so the
policy keeps tracking the latest pose rather than snapping to zero.
"""
frame = self._drain_latest()
now = time.time()
if frame is not None:
self._last_frame = frame
self._last_recv_t = now
self._warned_stale = False
if not self._got_first:
# Pre-fill the window so the first send is a full, coherent clip.
self._buf.extend([frame.copy() for _ in range(WINDOW)])
self._got_first = True
else:
self._buf.append(frame)
elif self._got_first:
# No fresh frame: repeat the most recent to keep the window moving.
self._buf.append(self._last_frame.copy())
if (
self.stale_after_s
and not self._warned_stale
and (now - self._last_recv_t) > self.stale_after_s
):
logger.warning(
"[SmplStream] no %s frame for %.2fs (holding last pose)",
SMPL_TOPIC,
now - self._last_recv_t,
)
self._warned_stale = True
if not self._got_first:
return np.zeros(SMPL_OBS_DIM, np.float32)
# Flatten to (720,): frames oldest->newest, joint-major within a frame
# [f0_j0_xyz, f0_j1_xyz, ..., f9_j23_xyz] — matches SmplMotion.window.
return np.concatenate(list(self._buf), dtype=np.float32).reshape(-1)
def close(self):
with contextlib.suppress(Exception):
self._poller.unregister(self._sock)
self._sock.close(0)
@@ -0,0 +1,147 @@
#!/usr/bin/env python
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Convert SMPL ``.npz`` motion clips into a LeRobotDataset for SONIC replay.
Each dataset frame's ``action`` is the 720-dim ``smpl.*`` window that the
``pico_headset`` teleoperator emits and ``SonicWholeBodyController`` reassembles
into ``encode_mode == 2``. So the resulting dataset can be pushed straight
through ``lerobot-replay`` to drive SONIC whole-body tracking with no headset:
lerobot-replay \
--robot.type=unitree_g1 \
--robot.controller=SonicWholeBodyController \
--dataset.repo_id=<user>/<clip> --dataset.episode=0
The 10-frame window is built exactly like the live ``SmplStream`` (oldest->newest,
the first frame repeated to pre-fill), so replayed actions match a live session.
Usage:
# One clip -> one-episode dataset
python -m lerobot.teleoperators.pico_headset.smpl_to_dataset \
--motion-file examples/unitree_g1/motions/walk_forward.npz \
--repo-id me/sonic_walk_forward
# Every clip in a dir -> one episode each
python -m lerobot.teleoperators.pico_headset.smpl_to_dataset \
--motion-dir examples/unitree_g1/motions --repo-id me/sonic_motions
"""
from __future__ import annotations
import argparse
from pathlib import Path
import numpy as np
from lerobot.teleoperators.pico_headset.smpl_constants import (
ACTION_DIM,
JOINT_DIM,
N_JOINTS,
ROOT_ACTION_DIM as ROOT_DIM,
ROOT_ACTION_PREFIX,
SMPL_ACTION_PREFIX,
SMPL_OBS_DIM,
WINDOW,
)
from lerobot.teleoperators.pico_headset.smpl_fk import canonicalize_smpl_joints, root_quats_from_aa
def _load_canonical_joints(path: str) -> tuple[np.ndarray, np.ndarray, float]:
"""Load an SMPL clip -> (canonical (T,24,3) joints, (T,4) root wxyz, fps)."""
data = np.load(path)
joints = data["smpl_joints"].astype(np.float32)
if joints.ndim != 3 or joints.shape[1:] != (N_JOINTS, JOINT_DIM):
raise ValueError(f"{path}: expected smpl_joints (T, 24, 3), got {joints.shape}")
t = joints.shape[0]
if "pose_aa" in data.files:
root_aa = data["pose_aa"].astype(np.float32)[:, :3]
joints = canonicalize_smpl_joints(joints, root_aa)
root_quat = root_quats_from_aa(root_aa) # (T, 4) wxyz, matches live stream
else:
# No global orient available: identity root (anchor falls back to standing).
root_quat = np.tile(np.array([1.0, 0.0, 0.0, 0.0], np.float32), (t, 1))
fps = float(data["fps"]) if "fps" in data.files else 50.0
return joints, root_quat, fps
def _windows(joints: np.ndarray) -> np.ndarray:
"""(T, 24, 3) -> (T, 720): rolling 10-frame window, matching SmplStream.
Window t = frames [t-9 .. t], clamped to 0 at the start (first frame repeated).
"""
t = joints.shape[0]
idx = np.clip(np.arange(t)[:, None] + np.arange(-WINDOW + 1, 1)[None, :], 0, t - 1)
return joints[idx].reshape(t, -1).astype(np.float32)
def _action_features() -> dict:
names = [f"{SMPL_ACTION_PREFIX}{i}" for i in range(SMPL_OBS_DIM)]
names += [f"{ROOT_ACTION_PREFIX}{i}" for i in range(ROOT_DIM)]
return {"action": {"dtype": "float32", "shape": (ACTION_DIM,), "names": names}}
def main() -> None:
p = argparse.ArgumentParser(description=__doc__)
src = p.add_mutually_exclusive_group(required=True)
src.add_argument("--motion-file", type=str, help="Single SMPL .npz clip")
src.add_argument("--motion-dir", type=str, help="Directory of .npz clips (one episode each)")
p.add_argument("--repo-id", required=True, help="Dataset repo id, e.g. user/name")
p.add_argument("--root", type=str, default=None, help="Local dataset root (default HF cache)")
p.add_argument("--fps", type=int, default=None, help="Override fps (default: clip fps)")
p.add_argument("--task", type=str, default="sonic whole-body SMPL replay")
args = p.parse_args()
from lerobot.datasets.lerobot_dataset import LeRobotDataset
if args.motion_dir:
clips = sorted(str(pth) for pth in Path(args.motion_dir).glob("*.npz"))
if not clips:
raise SystemExit(f"No .npz clips found in {args.motion_dir}")
else:
clips = [args.motion_file]
first_joints, first_root, first_fps = _load_canonical_joints(clips[0])
fps = args.fps or int(round(first_fps))
dataset = LeRobotDataset.create(
repo_id=args.repo_id,
fps=fps,
features=_action_features(),
root=args.root,
robot_type="unitree_g1",
use_videos=False,
)
for clip_i, clip in enumerate(clips):
if clip_i == 0:
joints, root_quat = first_joints, first_root
else:
joints, root_quat, _ = _load_canonical_joints(clip)
windows = _windows(joints) # (T, 720)
# action = [720 joint window | 4 root wxyz] per frame -> (T, 724)
actions = np.concatenate([windows, root_quat.astype(np.float32)], axis=1)
for a in actions:
dataset.add_frame({"action": a, "task": args.task})
dataset.save_episode()
print(f"[smpl_to_dataset] episode {clip_i}: {Path(clip).name} ({actions.shape[0]} frames)")
dataset.finalize()
print(f"[smpl_to_dataset] wrote {len(clips)} episode(s) to {dataset.root}")
if __name__ == "__main__":
main()
+14 -2
View File
@@ -60,8 +60,18 @@ def is_package_available(
# If the package can't be imported, it's not available
package_exists = False
else:
# For packages other than "torch", don't attempt the fallback and set as not available
package_exists = False
# The distribution may be published under a name that differs from the
# import name (e.g. ``onnxruntime`` imports from ``onnxruntime-gpu`` /
# ``onnxruntime-silicon``). Resolve the import name to its actual
# distribution(s) and read the version from there before giving up.
try:
dists = importlib.metadata.packages_distributions().get(import_name, [])
if dists:
package_version = importlib.metadata.version(dists[0])
else:
package_exists = False
except importlib.metadata.PackageNotFoundError:
package_exists = False
logging.debug(f"Detected {pkg_name} version: {package_version}")
if return_version:
return package_exists, package_version
@@ -123,6 +133,8 @@ _pyrealsense2_available = is_package_available("pyrealsense2") or is_package_ava
"pyrealsense2-macosx", import_name="pyrealsense2"
)
_zmq_available = is_package_available("pyzmq", import_name="zmq")
_onnxruntime_available = is_package_available("onnxruntime")
_onnx_available = is_package_available("onnx")
_hebi_available = is_package_available("hebi-py", import_name="hebi")
_teleop_available = is_package_available("teleop")
_placo_available = is_package_available("placo")