Omiretarget Kinematic Rl Eval

omiretarget-kinematic-rl-eval

OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction — Lujie Yang et al. (2025) (arXiv:2509.26633, 2025)

What this evaluates

Evaluates the kinematic feasibility and physical constraint satisfaction of retargeted humanoid motions, as well as the downstream reinforcement learning policy success rates for loco-manipulation and terrain interaction tasks.

Datasets

• OMOMO — total ?; splits: test (-1)
• In-house MoCap — total ?; splits: test (-1)
• LAFAN1 — total ?; splits: test (-1)

Metrics

• Penetration Duration (normalized) — range: [0, 1]
  • Ratio of time where the robot intersects with objects or terrain to the total trajectory length.
• Penetration Max Depth (cm) — range: other
  • Maximum intersection depth between the robot and objects/terrain during the trajectory.
• Foot Skating Duration (normalized) — range: [0, 1]
  • Ratio of time where the stance foot slides to the total desired foot sticking length.
• Foot Skating Max Velocity (cm/s) — range: other
  • Maximum sliding velocity of the stance foot during contact phases.
• Contact Preservation Duration (normalized) — range: [0, 1]
  • Ratio of time where contact is maintained (hand-object or foot/heel-terrain) to the desired contact length.
• Downstream RL Policy Success Rate (%) (primary) — range: percent
  • Percentage of episodes that meet the training termination criteria in simulation.

Input / output format

Input: Retargeted 3D motion trajectories (keyframes/poses) for a Unitree G1 humanoid, evaluated against reference human MoCap data and scene/object/terrain geometry.

Output: Quantitative scores for penetration duration, penetration max depth, foot skating duration, foot skating max velocity, contact preservation duration, and downstream RL policy success rate.

Scoring recipe

def compute_metrics(traj, scene_geom, contact_phases):
    # Penetration
    pen_dur = sum(1 for t in traj if intersects(traj[t], scene_geom)) / len(traj)
    pen_max = max(depth(traj[t], scene_geom) for t in traj if intersects(traj[t], scene_geom))
    # Foot Skating
    skate_dur = sum(1 for t in stance_foot if sliding_vel(traj[t]) > 0) / total_stance_time
    skate_max = max(sliding_vel(traj[t]) for t in stance_foot)
    # Contact Preservation
    cont_dur = sum(1 for t in contact_phases if is_contact_maintained(traj[t])) / len(contact_phases)
    # RL Success
    rl_success = count(successful_episodes) / total_episodes
    return pen_dur, pen_max, skate_dur, skate_max, cont_dur, rl_success

Common pitfalls

• GMR's high contact preservation score is misleading because scaling human hand keypoints to robot size often drives hands inside objects, causing substantial penetration.
• VideoMimic's soft collision cost conflicts with keypoint matching, degrading interaction preservation despite comparable terrain results.
• RL success rate is highly sensitive to kinematic artifacts; baselines with penetration/skating fail to train policies effectively without extensive reward engineering.

Evidence (verbatim from paper)

We evaluate the kinematic quality of retargeted motions on a Unitree G1 with three criteria: 1. Penetration: Measured by the time duration (normalized by the trajectory length) and maximum depth of intersections between the robot, objects, and terrain. 2. Foot Skating: Quantified by the time duration (normalized by the total desired foot sticking length) and maximum skating velocity of a stance foot. 3. Contact Preservation: Quantified by the time duration (normalized by the desired contact length).

Citation

@misc{yang2025omiretarget,
  title={OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction},
  author={Lujie Yang et al. (2025)},
  year={2025},
  note={arXiv:2509.26633}
}

• arXiv: 2509.26633