Add Unitree Go2 robot + spatial-memory navigation base

unitree_go2 robot: high-level sport-mode body-velocity control over DDS
(unitree_sdk2py) straight from the host — no companion computer, unlike
the G1's low-level ZMQ bridge. Actions x.vel/y.vel/theta.vel; observation
is planar odometry (rt/sportmodestate) + the built-in front camera via
VideoClient. SDK imported only in connect() so configs/features/tests
work without it; 20 tests via mocks. unitree_go2 pyproject extra.

lerobot.navigation package: BaseController protocol + StubBaseController +
SafeBaseController (velocity clamp, occupancy gate, keyframe watchdog,
e-stop latch) + RobotBaseController wrapping any Robot on the standard
REP-103 mobile-base contract, carrying the world<->body velocity and
odometry<->world pose frame math. Robot-agnostic: the SDK lives only in
the robot class. 29 tests, SDK/torch-free.

First step of consolidating the dyna360 DynaMem navigation stack into
lerobot; dyna360 is a source to copy from, not a runtime dependency.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/lerobot/navigation/README.md

@@ -0,0 +1,60 @@
+# `lerobot.navigation` — spatial-memory navigation
+
+Online spatio-semantic mapping (DynaMem-style), A* planning, obstacle
+avoidance and open-vocabulary goto/explore for LeRobot mobile bases.
+Ported from the dyna360 research stack; the physical robot layer lives in
+`lerobot.robots` (e.g. [`unitree_go2`](../robots/unitree_go2)).
+
+## Idea
+
+Drive any LeRobot `Robot` on the standard REP-103 mobile-base contract —
+body-velocity actions `x.vel`/`y.vel`/`theta.vel` and planar odometry
+`x.pos`/`y.pos`/`theta.pos` — from a spatial memory that is built and
+updated online from the robot's camera. With no prompt the base explores
+autonomously; given a text prompt it queries the map and navigates to the
+matching object, or explores to find it if it isn't there (or has moved).
+
+## Architecture
+
+The navigation layer talks to hardware only through LeRobot's own `Robot`
+interface, so it is robot-agnostic and carries no SDK dependency.
+
+```
+BaseController (protocol)         world-frame move()/pose() seam
+├── StubBaseController            kinematic integrator (sim, tests)
+├── RobotBaseController           wraps any Robot; world<->body +
+│                                 odometry<->world frame math
+└── SafeBaseController            velocity clamp, occupancy gate,
+                                  keyframe watchdog, e-stop latch
+```
+
+World frame is OpenCV (x right, y down, z forward); the base moves in the
+XZ plane. `RobotBaseController.feed_observation(obs)` updates pose from
+the observation the navigation loop already fetches (closed-loop
+odometry), avoiding an extra camera read; absent odometry it integrates
+open-loop so sim matches hardware.
+
+## Status (branch `feat/unitree-go2`)
+
+Implemented:
+- `base_controller.py` — the controller seam above. SDK/torch-free;
+  tests in `tests/navigation/test_base_controller.py`.
+
+Planned (porting from dyna360, everything lands here — dyna360 is a
+source to copy from, not a runtime dependency):
+- `geometry.py` — LingBot-Map streaming reconstruction runner
+  (`{points, local_points, conf, camera_poses}`), scale-anchored to
+  odometry. (Pi3X is not carried over.)
+- `voxel_map.py` — 5 cm `SegmentVoxelMap`, no point-cloud retention;
+  one SigLIP2 feature per segment.
+- `occupancy.py` — occupancy grid + A* + frontiers; `value_map.py`.
+- `features.py` (SigLIP2), `segmenter.py` (SAM2) — run offboard.
+- `agent.py`, `skills.py`, `dog_cli.py` — the interactive explore/query
+  REPL (the deliverable).
+
+## Target platform
+
+Unitree Go2 EDU, no companion computer: the workstation (single RTX 5090)
+talks DDS straight to the dog; geometry is monocular LingBot-Map from the
+built-in front camera, scale-anchored to sport-mode odometry; the map is
+5 cm voxels. See [`robots/unitree_go2`](../robots/unitree_go2).

src/lerobot/navigation/__init__.py

@@ -0,0 +1,42 @@
+#!/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.
+
+"""Spatial-memory navigation for LeRobot mobile bases.
+
+Online spatio-semantic mapping (DynaMem-style), A* planning, obstacle
+avoidance and open-vocabulary goto/explore, driving any LeRobot ``Robot``
+that exposes body-velocity actions and planar odometry. Ported from the
+dyna360 research stack; the physical robot layer lives in
+``lerobot.robots`` (e.g. ``unitree_go2``).
+"""
+
+from .base_controller import (
+    BaseController,
+    RobotBaseController,
+    SafeBaseController,
+    StubBaseController,
+    odometry_to_world_pose,
+    world_velocity_to_body,
+)
+
+__all__ = [
+    "BaseController",
+    "RobotBaseController",
+    "SafeBaseController",
+    "StubBaseController",
+    "odometry_to_world_pose",
+    "world_velocity_to_body",
+]

src/lerobot/navigation/base_controller.py

@@ -0,0 +1,389 @@
+#!/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.
+
+"""Base controller for spatial-memory navigation.
+
+The navigation/skills layer commands motion in a single **world frame**
+(OpenCV convention: x right, y down, z forward — the base lives in the XZ
+plane, y is gravity) and reads back an SE(3) pose. :class:`BaseController`
+is that seam. Three implementations:
+
+  - :class:`StubBaseController` — kinematic integrator, no hardware; sim +
+    unit tests.
+  - :class:`RobotBaseController` — drives any LeRobot :class:`Robot` whose
+    action space is body-frame velocities ``x.vel`` (forward, m/s),
+    ``y.vel`` (left, m/s), ``theta.vel`` (CCW yaw, rad/s) and whose
+    observation carries planar odometry ``x.pos``/``y.pos``/``theta.pos``
+    (REP-103: x forward, y left, yaw CCW). The Unitree Go2 satisfies this
+    out of the box; so would a LeKiwi base.
+  - :class:`SafeBaseController` — wraps any of the above with velocity
+    clamping, an optional occupancy gate, a keyframe watchdog and an
+    e-stop latch.
+
+All frame conversions between the world frame and a robot's body/odometry
+frame live in :func:`world_velocity_to_body` and
+:func:`odometry_to_world_pose`; nothing else needs to know the mapping.
+"""
+
+from __future__ import annotations
+
+import logging
+import math
+import time
+from abc import abstractmethod
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Protocol, runtime_checkable
+
+import numpy as np
+
+if TYPE_CHECKING:
+    from lerobot.robots import Robot
+
+logger = logging.getLogger(__name__)
+
+
+# --------------------------------------------------------------------- #
+# BaseController protocol
+# --------------------------------------------------------------------- #
+
+
+@runtime_checkable
+class BaseController(Protocol):
+    """Mobile-base interface used by the navigation/skills layer.
+
+    Velocities are in **world** frame XZ (m/s); ``yaw_rate`` is rad/s
+    about the world's −Y axis (turning around the up vector). ``pose``
+    is 4×4 SE(3) camera-to-world (OpenCV).
+    """
+
+    @abstractmethod
+    def move(self, vx: float, vz: float, yaw_rate: float = 0.0, dt: float = 0.05) -> None: ...
+
+    @abstractmethod
+    def stop(self) -> None: ...
+
+    @abstractmethod
+    def pose(self) -> np.ndarray: ...
+
+    @abstractmethod
+    def position(self) -> tuple[float, float, float]: ...
+
+
+# --------------------------------------------------------------------- #
+# Frame math (pure functions)
+# --------------------------------------------------------------------- #
+
+
+def world_velocity_to_body(
+    vx_world: float,
+    vz_world: float,
+    yaw_rate_rad_s: float,
+    heading_rad: float,
+) -> tuple[float, float, float]:
+    """World-frame velocity → body-frame ``(x.vel, y.vel, theta.vel)``.
+
+    Returns ``(vx_forward, vy_left, vyaw)`` in m/s, m/s, rad/s — the
+    action a REP-103 base expects. At heading ``h`` the body axes in the
+    world XZ plane are forward = (sin h, cos h), left = (−cos h, sin h)
+    (left = up × forward, up = −y). The navigation world's positive yaw
+    is clockwise about the up vector; a REP-103 base's ``theta.vel`` is
+    counter-clockwise, hence the sign flip.
+    """
+    s, c = math.sin(heading_rad), math.cos(heading_rad)
+    vx_fwd = vx_world * s + vz_world * c
+    vy_left = -vx_world * c + vz_world * s
+    return vx_fwd, vy_left, -yaw_rate_rad_s
+
+
+def odometry_to_world_pose(
+    x_fwd: float,
+    y_left: float,
+    yaw: float,
+    origin: tuple[float, float, float],
+) -> tuple[np.ndarray, float]:
+    """Planar odometry ``(x_fwd, y_left, yaw)`` → world pose + heading.
+
+    ``origin`` is the ``(x_fwd, y_left, yaw)`` sample captured when the
+    controller first saw odometry, so the run starts at identity
+    regardless of where the robot's odometry origin sits. The result is
+    the OpenCV world convention, planarized: height Y is 0 and only yaw
+    survives of the orientation — pitch/roll gait wobble is the camera's
+    concern, not the base's.
+
+    Odometry frame is REP-103 (x forward, y left, yaw CCW about z-up).
+    Mapping to OpenCV world: ``x_world = −y_odom``, ``z_world = x_odom``,
+    ``heading = −yaw``.
+    """
+    ox, oy, oyaw = origin
+    dx, dy = x_fwd - ox, y_left - oy
+    c0, s0 = math.cos(-oyaw), math.sin(-oyaw)
+    x_rel = c0 * dx - s0 * dy
+    y_rel = s0 * dx + c0 * dy
+    yaw_rel = yaw - oyaw
+
+    x_world, z_world = -y_rel, x_rel
+    heading = -yaw_rel
+
+    ch, sh = math.cos(heading), math.sin(heading)
+    pose = np.eye(4, dtype=np.float64)
+    pose[0, 0], pose[0, 2] = ch, sh
+    pose[2, 0], pose[2, 2] = -sh, ch
+    pose[0, 3], pose[2, 3] = x_world, z_world
+    return pose, heading
+
+
+def _heading_pose(x: float, z: float, heading: float) -> np.ndarray:
+    """Build a planar world pose from position + heading."""
+    c, s = math.cos(heading), math.sin(heading)
+    pose = np.eye(4, dtype=np.float64)
+    pose[0, 0], pose[0, 2] = c, s
+    pose[2, 0], pose[2, 2] = -s, c
+    pose[0, 3], pose[2, 3] = x, z
+    return pose
+
+
+# --------------------------------------------------------------------- #
+# Stub controller (kinematic, no hardware)
+# --------------------------------------------------------------------- #
+
+
+@dataclass
+class StubBaseController:
+    """Kinematic stub: integrates each ``move()`` into pose exactly.
+
+    No latency, slip or dynamics — for sim and skill-layer unit tests.
+    """
+
+    initial_pose: np.ndarray | None = None
+    max_lin_speed: float = 1.0
+    max_yaw_rate: float = 1.0
+
+    def __post_init__(self) -> None:
+        self._pose = (
+            np.asarray(self.initial_pose, dtype=np.float64).copy()
+            if self.initial_pose is not None
+            else np.eye(4, dtype=np.float64)
+        )
+        if self._pose.shape != (4, 4):
+            raise ValueError(f"initial_pose must be (4, 4); got {self._pose.shape}")
+        self._heading = 0.0
+        self._stopped = False
+
+    def move(self, vx: float, vz: float, yaw_rate: float = 0.0, dt: float = 0.05) -> None:
+        vx = float(np.clip(vx, -self.max_lin_speed, self.max_lin_speed))
+        vz = float(np.clip(vz, -self.max_lin_speed, self.max_lin_speed))
+        yaw_rate = float(np.clip(yaw_rate, -self.max_yaw_rate, self.max_yaw_rate))
+        if dt <= 0:
+            return
+        self._pose[0, 3] += vx * dt
+        self._pose[2, 3] += vz * dt
+        if yaw_rate != 0.0:
+            self._heading += yaw_rate * dt
+            self._pose = _heading_pose(self._pose[0, 3], self._pose[2, 3], self._heading)
+        self._stopped = False
+
+    def stop(self) -> None:
+        self._stopped = True
+
+    def pose(self) -> np.ndarray:
+        return self._pose.copy()
+
+    def position(self) -> tuple[float, float, float]:
+        p = self._pose[:3, 3]
+        return float(p[0]), float(p[1]), float(p[2])
+
+    @property
+    def is_stopped(self) -> bool:
+        return self._stopped
+
+
+# --------------------------------------------------------------------- #
+# Robot-backed controller
+# --------------------------------------------------------------------- #
+
+
+@dataclass(frozen=True)
+class RobotBaseControllerConfig:
+    """Behaviour knobs for :class:`RobotBaseController`."""
+
+    max_lin_speed: float = 0.6
+    """Hard cap on per-axis world linear velocity (m/s)."""
+
+    max_yaw_rate: float = 1.2
+    """Hard cap on yaw rate (rad/s)."""
+
+    pose_from_odometry: bool = True
+    """Report pose from the robot's odometry (closed-loop). When False,
+    integrate pose open-loop from commanded velocities."""
+
+
+class RobotBaseController(BaseController):
+    """:class:`BaseController` over any LeRobot :class:`Robot`.
+
+    The robot must accept body-velocity actions ``x.vel`` (forward),
+    ``y.vel`` (left), ``theta.vel`` (CCW yaw) and — for closed-loop pose
+    — report odometry ``x.pos``/``y.pos``/``theta.pos`` in its
+    observation. This is the standard REP-103 mobile-base contract, which
+    ``UnitreeGo2`` implements.
+
+    Pose is refreshed from observations the navigation loop already
+    fetches: call :meth:`feed_observation` each keyframe rather than
+    having the controller poll the robot (which would trigger an extra
+    camera read). Absent any fed observation, pose falls back to
+    open-loop integration so sim/dry-run behaves like the stub.
+    """
+
+    def __init__(self, robot: Robot, cfg: RobotBaseControllerConfig | None = None) -> None:
+        self.robot = robot
+        self.cfg = cfg or RobotBaseControllerConfig()
+        self._pose = np.eye(4, dtype=np.float64)
+        self._heading = 0.0
+        self._stopped = False
+        self._odom_origin: tuple[float, float, float] | None = None
+        self._have_odom = False
+
+    # ----- odometry feed --------------------------------------------------
+
+    def feed_observation(self, obs: dict) -> None:
+        """Update pose from an observation the nav loop already fetched."""
+        if not self.cfg.pose_from_odometry:
+            return
+        if not {"x.pos", "y.pos", "theta.pos"} <= obs.keys():
+            return
+        sample = (float(obs["x.pos"]), float(obs["y.pos"]), float(obs["theta.pos"]))
+        if self._odom_origin is None:
+            self._odom_origin = sample
+        self._pose, self._heading = odometry_to_world_pose(*sample, self._odom_origin)
+        self._have_odom = True
+
+    # ----- BaseController API --------------------------------------------
+
+    def move(self, vx: float, vz: float, yaw_rate: float = 0.0, dt: float = 0.05) -> None:
+        vx = float(np.clip(vx, -self.cfg.max_lin_speed, self.cfg.max_lin_speed))
+        vz = float(np.clip(vz, -self.cfg.max_lin_speed, self.cfg.max_lin_speed))
+        yaw_rate = float(np.clip(yaw_rate, -self.cfg.max_yaw_rate, self.cfg.max_yaw_rate))
+        if dt <= 0:
+            return
+
+        vx_fwd, vy_left, vyaw = world_velocity_to_body(vx, vz, yaw_rate, self._heading)
+        self.robot.send_action({"x.vel": vx_fwd, "y.vel": vy_left, "theta.vel": vyaw})
+
+        # Open-loop pose only when we have no odometry to trust.
+        if not (self.cfg.pose_from_odometry and self._have_odom):
+            self._pose[0, 3] += vx * dt
+            self._pose[2, 3] += vz * dt
+            if yaw_rate != 0.0:
+                self._heading += yaw_rate * dt
+            self._pose = _heading_pose(self._pose[0, 3], self._pose[2, 3], self._heading)
+        self._stopped = False
+
+    def stop(self) -> None:
+        self._stopped = True
+        try:
+            self.robot.send_action({"x.vel": 0.0, "y.vel": 0.0, "theta.vel": 0.0})
+        except Exception:
+            logger.exception("stop(): failed to send zero-velocity action")
+
+    def pose(self) -> np.ndarray:
+        return self._pose.copy()
+
+    def position(self) -> tuple[float, float, float]:
+        p = self._pose[:3, 3]
+        return float(p[0]), float(p[1]), float(p[2])
+
+    @property
+    def is_stopped(self) -> bool:
+        return self._stopped
+
+
+# --------------------------------------------------------------------- #
+# Safety wrapper
+# --------------------------------------------------------------------- #
+
+
+@dataclass
+class SafeBaseController(BaseController):
+    """Wrap any :class:`BaseController` with safety layers:
+
+    - **velocity clamp** on every ``move()``;
+    - **occupancy gate**: when ``occupancy_provider`` is set, predict
+      the next position and refuse (latch e-stop) if it lands in an
+      obstacle cell. The provider returns an object exposing
+      ``world_to_cell(x, z) -> (iz, ix)`` and an ``is_obstacle(iz, ix)
+      -> bool`` predicate; ``None`` means "no map yet, allow";
+    - **watchdog**: if no keyframe has been fed in
+      ``watchdog_timeout_s`` (caller ticks :meth:`feed_watchdog` per
+      map update), ``move()`` latches stop until :meth:`reset_watchdog`.
+    """
+
+    inner: BaseController
+    max_lin_speed: float = 0.6
+    max_yaw_rate: float = 1.2
+    occupancy_provider: object = None  # callable[[], grid | None] when set
+    watchdog_timeout_s: float = 2.0
+    e_stop_latched: bool = False
+    _last_keyframe_walltime: float = field(default_factory=time.monotonic, init=False)
+
+    def feed_watchdog(self) -> None:
+        self._last_keyframe_walltime = time.monotonic()
+
+    def reset_watchdog(self) -> None:
+        self.e_stop_latched = False
+        self._last_keyframe_walltime = time.monotonic()
+
+    def latch_estop(self, reason: str = "external") -> None:
+        logger.warning("SafeBaseController e-stop latched: %s", reason)
+        self.e_stop_latched = True
+        self.inner.stop()
+
+    def move(self, vx: float, vz: float, yaw_rate: float = 0.0, dt: float = 0.05) -> None:
+        if self.e_stop_latched:
+            return
+        if (time.monotonic() - self._last_keyframe_walltime) > self.watchdog_timeout_s:
+            self.latch_estop(f"watchdog: no keyframe in last {self.watchdog_timeout_s:.2f}s")
+            return
+
+        vx = float(np.clip(vx, -self.max_lin_speed, self.max_lin_speed))
+        vz = float(np.clip(vz, -self.max_lin_speed, self.max_lin_speed))
+        yaw_rate = float(np.clip(yaw_rate, -self.max_yaw_rate, self.max_yaw_rate))
+
+        if self.occupancy_provider is not None:
+            try:
+                grid = self.occupancy_provider()
+            except Exception:
+                logger.exception("occupancy_provider raised; refusing move")
+                return
+            if grid is not None and self._would_enter_obstacle(grid, vx, vz, dt):
+                self.latch_estop("about to enter obstacle cell")
+                return
+
+        self.inner.move(vx, vz, yaw_rate, dt)
+
+    def stop(self) -> None:
+        self.inner.stop()
+
+    def pose(self) -> np.ndarray:
+        return self.inner.pose()
+
+    def position(self) -> tuple[float, float, float]:
+        return self.inner.position()
+
+    def _would_enter_obstacle(self, grid, vx: float, vz: float, dt: float) -> bool:
+        pos = self.inner.position()
+        next_x = pos[0] + vx * dt
+        next_z = pos[2] + vz * dt
+        iz, ix = grid.world_to_cell(next_x, next_z)
+        return bool(grid.is_obstacle(iz, ix))

src/lerobot/robots/unitree_go2/__init__.py

@@ -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_unitree_go2 import UnitreeGo2Config
+from .unitree_go2 import UnitreeGo2
+
+__all__ = ["UnitreeGo2", "UnitreeGo2Config"]

src/lerobot/robots/unitree_go2/config_unitree_go2.py

@@ -0,0 +1,57 @@
+#!/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, field
+
+from lerobot.cameras import CameraConfig
+
+from ..config import RobotConfig
+
+
+@RobotConfig.register_subclass("unitree_go2")
+@dataclass
+class UnitreeGo2Config(RobotConfig):
+    """Configuration for the Unitree Go2 quadruped (EDU).
+
+    The host machine talks DDS directly to the dog over Ethernet/WiFi via
+    ``unitree_sdk2py`` — no onboard companion computer is required. Actions
+    are high-level sport-mode body velocities; observations are sport-mode
+    odometry plus the dog's built-in front camera.
+    """
+
+    # Network interface on the host that is wired/bridged to the Go2
+    # (the dog lives on 192.168.123.x when connected over Ethernet).
+    network_interface: str = "eth0"
+
+    # DDS domain id (0 for a stock Go2).
+    domain_id: int = 0
+
+    # Safety clamps applied in send_action() before commands reach the dog.
+    # The Go2 accepts far more (vx up to ~3.7 m/s) — keep indoor-sane defaults.
+    max_x_vel: float = 1.0  # m/s, body forward
+    max_y_vel: float = 0.5  # m/s, body left
+    max_theta_vel: float = 1.5  # rad/s, CCW about z-up
+
+    # Built-in front camera, served through the SDK VideoClient.
+    use_front_camera: bool = True
+    front_camera_width: int = 1280
+    front_camera_height: int = 720
+
+    # Send BalanceStand once on connect so the dog is ready to walk.
+    stand_on_connect: bool = True
+
+    # Additional external cameras (standard LeRobot camera configs).
+    cameras: dict[str, CameraConfig] = field(default_factory=dict)

src/lerobot/robots/unitree_go2/unitree_go2.py

@@ -0,0 +1,260 @@
+#!/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.
+
+"""Unitree Go2 quadruped (EDU) — high-level sport-mode integration.
+
+Unlike :class:`~lerobot.robots.unitree_g1.UnitreeG1` (low-level joint
+control at 250 Hz through an on-robot ZMQ bridge), the Go2 is driven with
+sport-mode **body velocity commands** at tens of Hz, which work fine over
+plain DDS from any Linux host on the dog's network — no bridge server, no
+onboard companion computer.
+
+Setup:
+  1. Connect the host to the Go2 via Ethernet (dog is on 192.168.123.x)
+     or put both on the same WiFi network.
+  2. ``pip install unitree_sdk2py`` (Linux only — rides on cyclonedds).
+  3. Find your interface name (``ip link``), then e.g.::
+
+       lerobot-teleoperate \
+           --robot.type=unitree_go2 \
+           --robot.network_interface=enp2s0 \
+           --teleop.type=gamepad
+
+Actions are body-frame velocities ``x.vel`` (forward, m/s), ``y.vel``
+(left, m/s), ``theta.vel`` (CCW yaw, rad/s) — the exact arguments of the
+SDK's ``SportClient.Move``. Observations are planar sport-mode odometry
+(``*.pos`` pose + ``*.vel`` body velocities) and the built-in front
+camera, plus any extra configured cameras.
+"""
+
+from __future__ import annotations
+
+import logging
+import threading
+from functools import cached_property
+from typing import Any
+
+import cv2
+import numpy as np
+
+from lerobot.cameras import make_cameras_from_configs
+from lerobot.types import RobotAction, RobotObservation
+from lerobot.utils.import_utils import require_package
+
+from ..robot import Robot
+from .config_unitree_go2 import UnitreeGo2Config
+
+logger = logging.getLogger(__name__)
+
+# DDS topic names follow Unitree SDK naming conventions
+SPORT_MODE_STATE_TOPIC = "rt/sportmodestate"
+
+
+class UnitreeGo2(Robot):
+    """LeRobot interface to a Unitree Go2 over unitree_sdk2py sport mode."""
+
+    config_class = UnitreeGo2Config
+    name = "unitree_go2"
+
+    def __init__(self, config: UnitreeGo2Config):
+        super().__init__(config)
+        self.config = config
+
+        self._cameras = make_cameras_from_configs(config.cameras)
+
+        # SDK handles — populated in connect(); the SDK import lives there
+        # too so that configs, features and tests work on SDK-less hosts.
+        self._sport = None
+        self._video = None
+        self._state_subscriber = None
+
+        self._state_lock = threading.Lock()
+        self._latest_state = None  # last SportModeState_ message
+        self._connected = False
+
+    # ------------------------------------------------------------------ #
+    # Features
+    # ------------------------------------------------------------------ #
+
+    @cached_property
+    def _odom_ft(self) -> dict[str, type]:
+        return {
+            "x.pos": float,
+            "y.pos": float,
+            "theta.pos": float,
+            "x.vel": float,
+            "y.vel": float,
+            "theta.vel": float,
+        }
+
+    @property
+    def _cameras_ft(self) -> dict[str, tuple]:
+        ft: dict[str, tuple] = {}
+        if self.config.use_front_camera:
+            ft["front"] = (self.config.front_camera_height, self.config.front_camera_width, 3)
+        for name, cam in self._cameras.items():
+            ft[name] = (cam.height, cam.width, 3)
+        return ft
+
+    @property
+    def observation_features(self) -> dict:
+        return {**self._odom_ft, **self._cameras_ft}
+
+    @property
+    def action_features(self) -> dict:
+        return {"x.vel": float, "y.vel": float, "theta.vel": float}
+
+    # ------------------------------------------------------------------ #
+    # Lifecycle
+    # ------------------------------------------------------------------ #
+
+    @property
+    def is_connected(self) -> bool:
+        return self._connected
+
+    def connect(self, calibrate: bool = True) -> None:
+        if self._connected:
+            return
+        require_package("unitree-sdk2py", extra="unitree_go2", import_name="unitree_sdk2py")
+
+        from unitree_sdk2py.core.channel import ChannelFactoryInitialize, ChannelSubscriber
+        from unitree_sdk2py.go2.sport.sport_client import SportClient
+        from unitree_sdk2py.idl.unitree_go.msg.dds_ import SportModeState_
+
+        ChannelFactoryInitialize(self.config.domain_id, self.config.network_interface)
+
+        sport = SportClient()
+        sport.SetTimeout(5.0)
+        sport.Init()
+        self._sport = sport
+
+        subscriber = ChannelSubscriber(SPORT_MODE_STATE_TOPIC, SportModeState_)
+        subscriber.Init(self._on_sport_state, 10)
+        self._state_subscriber = subscriber
+
+        if self.config.use_front_camera:
+            from unitree_sdk2py.go2.video.video_client import VideoClient
+
+            video = VideoClient()
+            video.SetTimeout(3.0)
+            video.Init()
+            self._video = video
+
+        for cam in self._cameras.values():
+            cam.connect()
+
+        if self.config.stand_on_connect:
+            self._sport.BalanceStand()
+
+        self._connected = True
+        self.configure()
+        logger.info(
+            "%s connected (iface=%s, domain=%d)",
+            self,
+            self.config.network_interface,
+            self.config.domain_id,
+        )
+
+    def disconnect(self) -> None:
+        if self._sport is not None:
+            try:
+                self._sport.StopMove()
+            except Exception:
+                logger.exception("StopMove on disconnect failed")
+        for cam in self._cameras.values():
+            try:
+                cam.disconnect()
+            except Exception:
+                logger.exception("camera disconnect failed")
+        self._sport = None
+        self._video = None
+        self._state_subscriber = None
+        self._connected = False
+
+    # Sport mode needs no calibration.
+    @property
+    def is_calibrated(self) -> bool:
+        return True
+
+    def calibrate(self) -> None:
+        pass
+
+    def configure(self) -> None:
+        pass
+
+    # ------------------------------------------------------------------ #
+    # I/O
+    # ------------------------------------------------------------------ #
+
+    def get_observation(self) -> RobotObservation:
+        if not self._connected:
+            raise ConnectionError(f"{self} is not connected.")
+
+        obs: dict[str, Any] = dict.fromkeys(self._odom_ft, 0.0)
+        with self._state_lock:
+            state = self._latest_state
+        if state is not None:
+            obs["x.pos"] = float(state.position[0])
+            obs["y.pos"] = float(state.position[1])
+            obs["theta.pos"] = float(state.imu_state.rpy[2])
+            obs["x.vel"] = float(state.velocity[0])
+            obs["y.vel"] = float(state.velocity[1])
+            obs["theta.vel"] = float(state.yaw_speed)
+
+        if self.config.use_front_camera:
+            obs["front"] = self._read_front_camera()
+
+        for name, cam in self._cameras.items():
+            obs[name] = cam.async_read()
+
+        return obs
+
+    def send_action(self, action: RobotAction) -> RobotAction:
+        if not self._connected:
+            raise ConnectionError(f"{self} is not connected.")
+
+        vx = float(np.clip(action.get("x.vel", 0.0), -self.config.max_x_vel, self.config.max_x_vel))
+        vy = float(np.clip(action.get("y.vel", 0.0), -self.config.max_y_vel, self.config.max_y_vel))
+        vyaw = float(
+            np.clip(action.get("theta.vel", 0.0), -self.config.max_theta_vel, self.config.max_theta_vel)
+        )
+
+        self._sport.Move(vx, vy, vyaw)
+        return {"x.vel": vx, "y.vel": vy, "theta.vel": vyaw}
+
+    # ------------------------------------------------------------------ #
+    # Internals
+    # ------------------------------------------------------------------ #
+
+    def _on_sport_state(self, msg) -> None:
+        with self._state_lock:
+            self._latest_state = msg
+
+    def _read_front_camera(self) -> np.ndarray:
+        """Fetch one frame from the built-in front camera (RGB, HxWx3)."""
+        h, w = self.config.front_camera_height, self.config.front_camera_width
+        code, data = self._video.GetImageSample()
+        if code != 0 or data is None:
+            logger.warning("front camera GetImageSample failed (code=%s)", code)
+            return np.zeros((h, w, 3), dtype=np.uint8)
+        frame = cv2.imdecode(np.frombuffer(bytes(data), dtype=np.uint8), cv2.IMREAD_COLOR)
+        if frame is None:
+            logger.warning("front camera frame failed to decode")
+            return np.zeros((h, w, 3), dtype=np.uint8)
+        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        if frame.shape[:2] != (h, w):
+            frame = cv2.resize(frame, (w, h), interpolation=cv2.INTER_AREA)
+        return frame