Example 3: Homie — Mixed Motion and Disturbances (Unitree G1 / H1)

Homie mixes velocity tracking, squatting (height control), standing, and upper-body random disturbances. The implementation follows the OpenHomie reference (HomieRL/legged_gym): reward set, command sampling, and curriculum.

Main task ids:

  • Mjlab-Homie-Unitree-G1: the OpenHomie robot (29 dof, 12 lower-body actions). The waist defaults to locked: waist_roll/pitch are PD-held at the default pose and excluded from the upper-body disturbance, matching OpenHomie’s 27-dof URDF (which welds those two joints) and real-robot deployment.

  • Mjlab-Homie-Unitree-H1: the H1 port (19 dof, 10 lower-body actions).

  • Mjlab-Homie-Unitree-H1-with_hands: H1 with Robotiq 2F85 grippers (policy-free random gripper motion and randomized hand payload).

G1 variants (identical observation/action interface; checkpoints load interchangeably):

  • Mjlab-Homie-Unitree-G1-free_waist: all three waist joints join the upper-body disturbance (a strict superset of the original distribution).

  • Mjlab-Homie-Unitree-G1-with_dex3 / -with_inspire: Unitree Dex3 / Inspire RH56 hand models mounted as inertial attachments with a randomized held-object payload. The base task’s wrist-payload randomization covers these hand masses, so one checkpoint serves bare wrists and both hand models; these variants are primarily for play/eval with real hand geometry.

  • Mjlab-Homie-Unitree-G1-mjlab_gains: ablation with mjlab’s first-principles actuator gains; sim-only.

The core idea: reduce the policy action space to the lower body, and treat the upper body (and optional grippers) as smooth, time-varying disturbances.

Task registration

Path: src/mjlab_homierl/__init__.py

The task ids above are registered via mjlab.tasks.registry.register_mjlab_task. The play env is the same config with play=True (critic observations, rewards, and curriculum stripped; upper-body motion at full range).

Three-mode command sampling

Path: src/mjlab_homierl/mdp/velocity_command.py

Commands resample every 4 s. Each environment draws one of three mutually exclusive modes (the OpenHomie scheme):

  • squat (p = 1/3): zero twist, random relative-height target;

  • walk (p = 1/2): random twist (x ∈ [-0.8, 1.2], y ∈ [-0.5, 0.5], yaw ∈ [-0.8, 0.8]), standing-height target;

  • stand (p = 1/6): zero twist, standing-height target.

The twist command samples the mode and exposes it via its mode attribute; the height command (RelativeHeightCommand, pelvis height above the lowest foot site) couples to it. Both commands must share the same resampling interval, and twist must precede height in the commands dict.

Height ranges scale with the robot: G1 stands at 0.78 m and squats down to 0.28 m; H1 stands at 0.98 m and squats to 0.4 m. A set of “standing only” reward terms (hip/ankle deviation, feet_parallel, stand_still, …) are gated on the commanded height being near the standing height.

Reward set

Paths: src/mjlab_homierl/homie_env_cfg.py (weights), src/mjlab_homierl/mdp/rewards.py (implementations)

Terms and weights follow OpenHomie’s G1 config: split x/y velocity tracking (1.5 / 1.0), yaw tracking (2.0, σ²=0.25), height tracking exp(-4|err|) (2.0), hip/ankle deviation (-0.2 / -0.5), knee-driven squatting (-0.75), the full joint regularization suite (torques, power, velocity, acceleration, soft limits), and the feet terms (air time, no-fly, clearance, slip, stumble, contact forces, contact momentum, sole parallelism, feet parallel, lateral distances).

Intentional deviations from OpenHomie:

  • IsaacGym’s torso-contact termination is replaced by contact penalties plus a torso-contact termination on G1, with additional self-collision and hip/knee ground-contact penalties.

  • The fall-over termination threshold matches the original (asin(0.8) ≈ 53°).

Upper-body disturbance and curriculum

Paths: src/mjlab_homierl/mdp/actions.py, mdp/curriculums.py

UpperBodyPoseAction contributes zero policy dimensions. Every 1 s (a global interval event), upper-body goal poses are resampled for all environments: amplitudes come from a truncated-exponential transform of the curriculum ratio (heavily biased toward small motions early on), the direction is a fair coin between each joint’s lower/upper hard limit (so amplitudes are proportional to joint range), and the target is reached by linear interpolation over one interval.

Curriculum advancement matches OpenHomie: the check runs only when common_step_counter is a multiple of the max episode length; if the episode-average raw forward-velocity tracking reward is ≥ 0.8, the global ratio increases by 0.05.

Domain randomization

All randomization uses mjlab-native dr.* events: PD gains ×[0.9, 1.1] (per reset), link masses ×[0.8, 1.2], torso payload +[-1, +5] kg (see the env cfg comment for the survey behind this deviation), CoM offset, encoder bias, foot friction, a global horizontal push every 4 s (Δv ≤ 0.5 m/s), and randomized joint poses / root velocities at reset. OpenHomie’s per-step torque injection has no mjlab equivalent and is approximated by PD-gain randomization and encoder bias.

HIM-PPO

Path: src/mjlab_homierl/rl/himppo/

Hyperparameters and network sizes (actor/critic hidden dims 512-256-256, estimator latent 32, prototype 64, sinkhorn contrastive loss) follow OpenHomie. The left/right mirror maps for symmetry augmentation are derived from joint names (left_*/right_* pairing; joints whose names contain yaw/roll flip sign), so G1 and H1 share one implementation.

Termination steps feed the estimator the pre-reset critic observation via a recorder term, matching OpenHomie’s runner behavior.