feat: add Microduck robot and integrate it into humanoid walk task

configs/task/microduck-walk-flat/motrix.fastsac.yaml

@@ -0,0 +1,19 @@
+# @package _global_
+# Task microduck-walk-flat: motrix.fastsac shared training recipe.
+defaults:
+  - /algo_base@algo: motrix.fastsac
+  - _self_
+task:
+  env: microduck-walk-flat
+  rllib: motrix
+  algo: fastsac
+num_envs: 2048
+play_num_envs: 16
+seed: 1
+checkpoint:
+  interval: 1000
+algo:
+  agent:
+    alpha_init: 0.01
+    target_entropy_ratio: -0.1
+    num_updates: 4

configs/task/microduck-walk-rough/motrix.fastsac.yaml

@@ -0,0 +1,19 @@
+# @package _global_
+# Task microduck-walk-rough: motrix.fastsac shared training recipe.
+defaults:
+  - /algo_base@algo: motrix.fastsac
+  - _self_
+task:
+  env: microduck-walk-rough
+  rllib: motrix
+  algo: fastsac
+num_envs: 2048
+play_num_envs: 16
+seed: 1
+checkpoint:
+  interval: 1000
+algo:
+  agent:
+    alpha_init: 0.01
+    target_entropy_ratio: -0.1
+    num_updates: 4

docs/scripts/generate_robot_docs.py

@@ -38,6 +38,7 @@
     "go1": {"kind": "quadruped", "lookat": (0.0, 0.0, 0.3), "distance": 1.4},
     "go2": {"kind": "quadruped", "lookat": (0.0, 0.0, 0.3), "distance": 1.4},
     "k1": {"kind": "humanoid", "lookat": (0.0, 0.0, 0.8), "distance": 2.1},
+    "microduck": {"kind": "humanoid", "lookat": (0.0, 0.0, 0.15), "distance": 0.8},
 }
 
 _DOC_CONFIGS = {

docs/source/en/user_guide/envs/index.md

@@ -66,6 +66,8 @@ uv run scripts/view.py env=<env-id>
 | <img src="../../_static/images/poster/k1-walk-flat.jpg" alt="k1-walk-flat" width="240"> | `k1-walk-flat` | Track walking commands with Booster K1 on flat ground. | `motrix.fastsac` |
 | <img src="../../_static/images/poster/k1-walk-rough.jpg" alt="k1-walk-rough" width="240"> | `k1-walk-rough` | Track walking commands with Booster K1 over uneven terrain. | `motrix.fastsac` |
 | <img src="../../_static/images/poster/k1-wbt-freekick.jpg" alt="k1-wbt-freekick" width="240"> | `k1-wbt-freekick` | Track a free-kick reference motion with Booster K1. | `motrix.fastsac` |
+| <img src="../../_static/images/poster/microduck-walk-flat.jpg" alt="microduck-walk-flat" width="240"> | `microduck-walk-flat` | Track walking commands with Microduck on flat ground. | `motrix.fastsac` |
+| <img src="../../_static/images/poster/microduck-walk-rough.jpg" alt="microduck-walk-rough" width="240"> | `microduck-walk-rough` | Track walking commands with Microduck over uneven terrain. | `motrix.fastsac` |
 | <img src="../../_static/images/poster/peg-insert.jpg" alt="peg-insert" width="240"> | `peg-insert` | Control RM65 to grasp, align, and insert a peg into a socket. | `motrix.fastsac`, `skrl.ppo` |
 | <img src="../../_static/images/poster/pendulum.jpg" alt="pendulum" width="240"> | `pendulum` | Apply joint torque to swing up and balance a pendulum. | `rslrl.ppo`, `skrl.ppo` |
 | <img src="../../_static/images/poster/point_mass.jpg" alt="point_mass" width="240"> | `point_mass` | Move a two-dimensional point mass to a random target. | `skrl.ppo` |

docs/source/en/user_guide/robots.md

@@ -8,15 +8,15 @@ be composed into different scenes and tasks. The table below is generated direct
 <!-- ROBOT_TABLE_START -->
 
 <!-- This table is generated; do not edit this block manually. -->
-
-| Screenshot                                                                   | Registry name | Configuration class | Type      | Model format | DoF |
-| ---------------------------------------------------------------------------- | ------------- | ------------------- | --------- | ------------ | --- |
-| <img src="../_static/images/robots/anymal_c.png" alt="anymal_c" width="180"> | `anymal_c`    | `AnymalC`           | Quadruped | MJCF         | 12  |
-| <img src="../_static/images/robots/dex-evt.png" alt="dex-evt" width="180">   | `dex-evt`     | `DexEvt`            | Humanoid  | URDF         | 23  |
-| <img src="../_static/images/robots/g1-29dof.png" alt="g1-29dof" width="180"> | `g1-29dof`    | `UnitreeG129Dof`    | Humanoid  | MJCF         | 29  |
-| <img src="../_static/images/robots/go1.png" alt="go1" width="180">           | `go1`         | `UnitreeGo1Robot`   | Quadruped | MJCF         | 12  |
-| <img src="../_static/images/robots/go2.png" alt="go2" width="180">           | `go2`         | `UnitreeGo2Robot`   | Quadruped | MJCF         | 12  |
-| <img src="../_static/images/robots/k1.png" alt="k1" width="180">             | `k1`          | `BoosterK1`         | Humanoid  | MJCF         | 22  |
+| Screenshot | Registry name | Configuration class | Type | Model format | DoF |
+| --- | --- | --- | --- | --- | --- |
+| <img src="../_static/images/robots/anymal_c.png" alt="anymal_c" width="180"> | `anymal_c` | `AnymalC` | Quadruped | MJCF | 12 |
+| <img src="../_static/images/robots/dex-evt.png" alt="dex-evt" width="180"> | `dex-evt` | `DexEvt` | Humanoid | URDF | 23 |
+| <img src="../_static/images/robots/g1-29dof.png" alt="g1-29dof" width="180"> | `g1-29dof` | `UnitreeG129Dof` | Humanoid | MJCF | 29 |
+| <img src="../_static/images/robots/go1.png" alt="go1" width="180"> | `go1` | `UnitreeGo1Robot` | Quadruped | MJCF | 12 |
+| <img src="../_static/images/robots/go2.png" alt="go2" width="180"> | `go2` | `UnitreeGo2Robot` | Quadruped | MJCF | 12 |
+| <img src="../_static/images/robots/k1.png" alt="k1" width="180"> | `k1` | `BoosterK1` | Humanoid | MJCF | 22 |
+| <img src="../_static/images/robots/microduck.png" alt="microduck" width="180"> | `microduck` | `Microduck` | Humanoid | MJCF | 14 |
 
 <!-- ROBOT_TABLE_END -->
 

docs/source/zh_CN/user_guide/envs/index.md

@@ -66,6 +66,8 @@ uv run scripts/view.py env=<env-id>
 | <img src="../../_static/images/poster/k1-walk-flat.jpg" alt="k1-walk-flat" width="240"> | `k1-walk-flat` | 控制 Booster K1 在平地上跟踪行走指令。 | `motrix.fastsac` |
 | <img src="../../_static/images/poster/k1-walk-rough.jpg" alt="k1-walk-rough" width="240"> | `k1-walk-rough` | 控制 Booster K1 在起伏地形上跟踪行走指令。 | `motrix.fastsac` |
 | <img src="../../_static/images/poster/k1-wbt-freekick.jpg" alt="k1-wbt-freekick" width="240"> | `k1-wbt-freekick` | 让 Booster K1 跟踪任意球射门参考动作。 | `motrix.fastsac` |
+| <img src="../../_static/images/poster/microduck-walk-flat.jpg" alt="microduck-walk-flat" width="240"> | `microduck-walk-flat` | 控制 Microduck 小型双足机器人在平地上跟踪行走指令。 | `motrix.fastsac` |
+| <img src="../../_static/images/poster/microduck-walk-rough.jpg" alt="microduck-walk-rough" width="240"> | `microduck-walk-rough` | 控制 Microduck 小型双足机器人在起伏地形上跟踪行走指令。 | `motrix.fastsac` |
 | <img src="../../_static/images/poster/peg-insert.jpg" alt="peg-insert" width="240"> | `peg-insert` | 控制 RM65 抓取插销、对准插座并完成插入。 | `motrix.fastsac`, `skrl.ppo` |
 | <img src="../../_static/images/poster/pendulum.jpg" alt="pendulum" width="240"> | `pendulum` | 施加关节力矩使单摆摆起并保持直立。 | `rslrl.ppo`, `skrl.ppo` |
 | <img src="../../_static/images/poster/point_mass.jpg" alt="point_mass" width="240"> | `point_mass` | 控制二维质点移动到随机目标位置。 | `skrl.ppo` |

docs/source/zh_CN/user_guide/robots.md

@@ -8,15 +8,15 @@ scene 和 task 中。下表直接根据 registry 和 robot 配置生成。
 <!-- ROBOT_TABLE_START -->
 
 <!-- This table is generated; do not edit this block manually. -->
-
-| 截图                                                                         | Registry 名称 | 配置类            | 类型       | 模型格式 | 自由度 |
-| ---------------------------------------------------------------------------- | ------------- | ----------------- | ---------- | -------- | ------ |
-| <img src="../_static/images/robots/anymal_c.png" alt="anymal_c" width="180"> | `anymal_c`    | `AnymalC`         | 四足机器人 | MJCF     | 12     |
-| <img src="../_static/images/robots/dex-evt.png" alt="dex-evt" width="180">   | `dex-evt`     | `DexEvt`          | 人形机器人 | URDF     | 23     |
-| <img src="../_static/images/robots/g1-29dof.png" alt="g1-29dof" width="180"> | `g1-29dof`    | `UnitreeG129Dof`  | 人形机器人 | MJCF     | 29     |
-| <img src="../_static/images/robots/go1.png" alt="go1" width="180">           | `go1`         | `UnitreeGo1Robot` | 四足机器人 | MJCF     | 12     |
-| <img src="../_static/images/robots/go2.png" alt="go2" width="180">           | `go2`         | `UnitreeGo2Robot` | 四足机器人 | MJCF     | 12     |
-| <img src="../_static/images/robots/k1.png" alt="k1" width="180">             | `k1`          | `BoosterK1`       | 人形机器人 | MJCF     | 22     |
+| 截图 | Registry 名称 | 配置类 | 类型 | 模型格式 | 自由度 |
+| --- | --- | --- | --- | --- | --- |
+| <img src="../_static/images/robots/anymal_c.png" alt="anymal_c" width="180"> | `anymal_c` | `AnymalC` | 四足机器人 | MJCF | 12 |
+| <img src="../_static/images/robots/dex-evt.png" alt="dex-evt" width="180"> | `dex-evt` | `DexEvt` | 人形机器人 | URDF | 23 |
+| <img src="../_static/images/robots/g1-29dof.png" alt="g1-29dof" width="180"> | `g1-29dof` | `UnitreeG129Dof` | 人形机器人 | MJCF | 29 |
+| <img src="../_static/images/robots/go1.png" alt="go1" width="180"> | `go1` | `UnitreeGo1Robot` | 四足机器人 | MJCF | 12 |
+| <img src="../_static/images/robots/go2.png" alt="go2" width="180"> | `go2` | `UnitreeGo2Robot` | 四足机器人 | MJCF | 12 |
+| <img src="../_static/images/robots/k1.png" alt="k1" width="180"> | `k1` | `BoosterK1` | 人形机器人 | MJCF | 22 |
+| <img src="../_static/images/robots/microduck.png" alt="microduck" width="180"> | `microduck` | `Microduck` | 人形机器人 | MJCF | 14 |
 
 <!-- ROBOT_TABLE_END -->
 

motrix_envs/src/motrix_envs/locomotion/humanoid/__init__.py

@@ -3,4 +3,4 @@
 
 """Robot-agnostic humanoid locomotion environments and configuration schemas."""
 
-from . import dex_evt, g1, k1, walk_np  # noqa: F401 register envs and expose base task
+from . import dex_evt, g1, k1, microduck, walk_np  # noqa: F401 register envs and expose base task

motrix_envs/src/motrix_envs/locomotion/humanoid/microduck.py

@@ -0,0 +1,123 @@
+# Copyright Motphys Technology Co., Ltd. 2025, 2026
+# SPDX-License-Identifier: Apache-2.0
+
+"""Pollen Robotics Microduck humanoid velocity-tracking presets and registration."""
+
+from dataclasses import replace
+
+from motrix_env_core import registry
+from motrix_env_core.base import SimCfg
+from motrix_env_core.config.scene import HFieldTerrainCfg, SystemCameraCfg
+from motrix_envs.config.scene import StandardSceneObjsCfg
+from motrix_envs.locomotion.humanoid import cfg as humanoid_cfg
+from motrix_envs.locomotion.humanoid.walk_np import HumanoidVelocityTrackingEnv
+from motrix_envs.robot import Microduck
+
+
+def _make_microduck_robot() -> Microduck:
+    return Microduck()
+
+
+# Pinned Microduck termination collision inventory, owned by the Microduck walk task.
+_MICRODUCK_TERMINATION_GEOMS = ("trunk_collision",)
+
+
+@registry.envcfg("microduck-walk-flat")
+def make_microduck_walk_flat_cfg() -> humanoid_cfg.HumanoidVelocityTrackingEnvCfg:
+    """Track walking commands with Microduck on flat ground.
+
+    zh_CN: 控制 Microduck 小型双足机器人在平地上跟踪行走指令。
+    """
+
+    return humanoid_cfg.HumanoidVelocityTrackingEnvCfg(
+        scene=humanoid_cfg.HumanoidWalkSceneCfg(
+            # Microduck is a ~25 cm robot; frame the lead env much closer than
+            # the full-size humanoid default camera (grid center at z=0.75).
+            system_camera=SystemCameraCfg(
+                lookat=(0.0, 0.0, 0.12),
+                distance=0.35,
+                elevation=-20.0,
+                azimuth=180.0,
+            ),
+            objs=StandardSceneObjsCfg(robot=_make_microduck_robot()),
+        ),
+        control_config=humanoid_cfg.ControlCfg(action_scale=0.5),
+        commands=humanoid_cfg.CommandsCfg(
+            vel_limit=[
+                [-1.0, -1.0, -1.0],
+                [1.0, 1.0, 1.0],
+            ],
+        ),
+        gait=humanoid_cfg.GaitCfg(
+            period=0.5,
+            swing_height=0.04,
+            feet_phase_sigma=0.002,
+        ),
+        reward_config=humanoid_cfg.RewardCfg(
+            scales=humanoid_cfg.RewardScales(
+                tracking_lin_vel=10.0,
+                tracking_ang_vel=3.0,
+                penalty_action_rate=-0.5,
+                feet_phase=8.0,
+            ),
+            tracking_sigma=0.15,
+            close_feet_threshold=0.05,
+            pose_weights={
+                "left_hip_yaw": 5.0,
+                "left_hip_roll": 1.0,
+                "left_hip_pitch": 0.01,
+                "left_knee": 0.01,
+                "left_ankle": 5.0,
+                "neck_pitch": 50.0,
+                "head_pitch": 50.0,
+                "head_yaw": 50.0,
+                "head_roll": 50.0,
+                "right_hip_yaw": 5.0,
+                "right_hip_roll": 1.0,
+                "right_hip_pitch": 0.01,
+                "right_knee": 0.01,
+                "right_ankle": 5.0,
+            },
+        ),
+        asset=humanoid_cfg.AssetCfg(
+            foot_height_site_names=("left_foot", "right_foot"),
+            ground_geom_name="floor",
+            terminate_contact_geom_names=_MICRODUCK_TERMINATION_GEOMS,
+        ),
+        sim=SimCfg(dt=0.005, solver_iterations=6, solver_tolerance=1e-4),
+        spawn_xy_range=0.0,
+    )
+
+
+@registry.envcfg("microduck-walk-rough")
+def make_microduck_walk_rough_cfg() -> humanoid_cfg.HumanoidVelocityTrackingEnvCfg:
+    """Track walking commands with Microduck over uneven terrain.
+
+    zh_CN: 控制 Microduck 小型双足机器人在起伏地形上跟踪行走指令。
+    """
+
+    return replace(
+        make_microduck_walk_flat_cfg(),
+        scene=humanoid_cfg.HumanoidWalkSceneCfg(
+            assets=humanoid_cfg.TerrainSceneAssetsCfg(),
+            system_camera=SystemCameraCfg(
+                lookat=(0.0, 0.0, 0.12),
+                distance=0.35,
+                elevation=-20.0,
+                azimuth=180.0,
+            ),
+            objs=StandardSceneObjsCfg(
+                floor=HFieldTerrainCfg(
+                    hfield="terrain",
+                    material="mat_ground",
+                ),
+                robot=_make_microduck_robot(),
+            ),
+        ),
+        spawn_xy_range=4.0,
+        render_spacing=0.0,
+    )
+
+
+registry.env("microduck-walk-flat")(HumanoidVelocityTrackingEnv)
+registry.env("microduck-walk-rough")(HumanoidVelocityTrackingEnv)

motrix_envs/src/motrix_envs/locomotion/humanoid/walk_np.py

@@ -171,6 +171,11 @@ def _init_buffers(self) -> None:
         self._phase_dt = 2.0 * np.pi * self._gait_freq * cfg.ctrl_dt
         self._termination_geoms = self._resolve_termination_geoms()
         self._num_termination_pairs = len(self._termination_geoms)
+        # Foot-link gravity direction at the default pose, captured on first
+        # reset. Foot orientation penalties are measured against this
+        # reference so robots whose ankle-link frames are not world-aligned
+        # (e.g. onshape-to-robot exports) are scored correctly.
+        self._default_foot_gravity: np.ndarray | None = None
 
         cur = cfg.curriculum
         self._penalty_terms = set(cur.penalty_terms)
@@ -392,6 +397,9 @@ def reset(self, env_ids: np.ndarray) -> dict:
         self._reset_joint_velocity[env_ids] = 0.0
         self._reset_program.execute(env_ids)
         self.sim_data.execute(row_ids)
+        if self._default_foot_gravity is None:
+            foot_quat = self.sim_data["foot_quat"][row_ids[0]]
+            self._default_foot_gravity = quaternion.rotate_inverse(foot_quat, self.gravity_vec)
 
         info = {
             "current_actions": np.zeros((num_reset, self._num_action), dtype=np.float32),
@@ -462,8 +470,12 @@ def _r_close_feet(self, q: StateQuantities, threshold: float) -> np.ndarray:
         return (distance < threshold).astype(np.float32)
 
     def _r_feet_ori(self, q: StateQuantities) -> np.ndarray:
+        # Magnitude of the cross product between the current foot-frame gravity
+        # and its default-pose reference. With a world-aligned default
+        # (gravity = (0, 0, -1)) this reduces exactly to sqrt(gx^2 + gy^2).
         total = np.zeros((q.foot_quat.shape[0],), dtype=np.float32)
         for foot_index in range(2):
             foot_gravity = quaternion.rotate_inverse(q.foot_quat[:, foot_index], self.gravity_vec)
-            total = total + np.sqrt(np.sum(np.square(foot_gravity[:, :2]), axis=1))
+            reference = self._default_foot_gravity[foot_index]
+            total = total + np.linalg.norm(np.cross(foot_gravity, reference), axis=1)
         return total

motrix_envs/src/motrix_envs/robot/__init__.py

@@ -6,6 +6,7 @@
 from motrix_envs.robot.booster import BoosterK1
 from motrix_envs.robot.dex_evt import DexEvt
 from motrix_envs.robot.humanoid import HumanoidRobotCfg
+from motrix_envs.robot.microduck import Microduck
 from motrix_envs.robot.quadruped import QuadrupedLegCfg, QuadrupedLegsCfg, QuadrupedRobotCfg
 from motrix_envs.robot.unitree import (
     UnitreeG129Dof,
@@ -19,12 +20,14 @@
 registry.robotcfg("go1")(UnitreeGo1Robot)
 registry.robotcfg("go2")(UnitreeGo2Robot)
 registry.robotcfg("k1")(BoosterK1)
+registry.robotcfg("microduck")(Microduck)
 
 __all__ = [
     "AnymalC",
     "BoosterK1",
     "DexEvt",
     "HumanoidRobotCfg",
+    "Microduck",
     "QuadrupedLegCfg",
     "QuadrupedLegsCfg",
     "QuadrupedRobotCfg",

motrix_envs/src/motrix_envs/robot/assets/microduck/README.md

@@ -0,0 +1,14 @@
+# Microduck robot model
+
+Ported from https://github.com/pollen-robotics/microduck_rl
+(`src/mjlab_microduck/robot/microduck/robot_walk.xml`, renamed to
+`microduck.xml`). Relative to upstream, a named termination capsule
+`trunk_collision` was added on `trunk_base` for the humanoid walk task's
+fall-termination contact query; everything else is unmodified.
+
+- Upstream code and MJCF: Apache License 2.0
+- 3D mesh files (`assets/*.stl`): Creative Commons BY-SA-NC
+  (non-commercial), per the upstream README
+
+The Microduck is a ~800 g, ~25 cm bipedal robot by Pollen Robotics with 14
+actuated joints (5 per leg, 4 neck/head).