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>

src/lerobot/robots/unitree_g1/config_unitree_g1.py

@@ -65,6 +65,35 @@ 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
 

src/lerobot/robots/unitree_g1/controllers/sonic_whole_body.py

@@ -20,6 +20,7 @@ from __future__ import annotations
 
 import logging
 import math
+from collections import deque
 from typing import TYPE_CHECKING
 
 import numpy as np
@@ -38,10 +39,12 @@ from lerobot.teleoperators.pico_headset.smpl_constants import (
     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 KEYBOARD_KEYS_FIELD, G1_29_JointIndex, lowstate_to_obs
+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,
@@ -132,6 +135,51 @@ def _extract_vr3_from_action(action: dict | None) -> tuple[np.ndarray, np.ndarra
     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}``.
 
@@ -234,6 +282,10 @@ class SonicWholeBodyController:
 
     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,
@@ -277,6 +329,13 @@ class SonicWholeBodyController:
         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.
@@ -445,6 +504,66 @@ class SonicWholeBodyController:
             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).
 
@@ -513,6 +632,20 @@ class SonicWholeBodyController:
         # 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)
@@ -575,6 +708,9 @@ class SonicWholeBodyController:
         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:

src/lerobot/robots/unitree_g1/g1_utils.py

@@ -104,6 +104,21 @@ REMOTE_KEYS = REMOTE_AXES + REMOTE_BUTTONS
 # 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)."""
@@ -181,6 +196,43 @@ def lowstate_to_obs(lowstate) -> dict:
     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:

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -40,6 +40,7 @@ from .g1_utils import (
     default_remote_input,
     lowstate_to_obs,
     make_locomotion_controller,
+    obs_to_wb34_state,
 )
 
 if TYPE_CHECKING or _unitree_sdk_available:
@@ -48,8 +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,
     )
@@ -59,6 +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
@@ -158,6 +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()
@@ -232,15 +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}
@@ -312,6 +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()
@@ -460,6 +548,31 @@ class UnitreeG1(Robot):
         # Motors + IMU + wireless remote (shared lowstate -> obs mapping)
         obs = lowstate_to_obs(lowstate)
 
+        # 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)
+
+        # 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
+
+        # 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():
             if getattr(cam, "use_rgb", True):
@@ -473,6 +586,8 @@ class UnitreeG1(Robot):
         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
@@ -507,10 +622,10 @@ class UnitreeG1(Robot):
         """Update controller input state from an incoming teleop action.
 
         Controller-agnostic: every value-carrying key is forwarded verbatim into
-        ``controller_input`` (joystick ``remote.*``, whole-body ``smpl.*``/``root.*``
-        from pico_headset, 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.
+        ``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
@@ -528,6 +643,44 @@ class UnitreeG1(Robot):
                 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:
         return True

src/lerobot/teleoperators/pico_headset/pico_publisher.py

@@ -195,7 +195,13 @@ def main() -> None:
     # 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.
-    calibrator = ThreePointCalibrator() if (xrt is not None and args.headset_source == "devices") else None
+    # 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(

src/lerobot/teleoperators/pico_headset/smpl_constants.py

@@ -59,3 +59,22 @@ 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"

src/lerobot/teleoperators/pico_headset/smpl_fk.py

@@ -394,8 +394,20 @@ class ThreePointCalibrator:
     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)
@@ -436,13 +448,15 @@ class ThreePointCalibrator:
         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()
-        neck_inv = self._neck_quat_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 = neck_inv.apply(pos[k])
-            corrected_rot = neck_inv * R.from_quat(orn[k], scalar_first=True)
+            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).
@@ -461,12 +475,13 @@ class ThreePointCalibrator:
         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] = neck_inv.apply(pos[k]) - self._wrist_pos_offset[k]
-            corrected_rot = neck_inv * R.from_quat(orn[k], scalar_first=True)
+            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.