Issue with G1 + Inspire Dexterous Hands

[francesco-dorati]

Environment

• Robot: Unitree G1 with Inspire Dexterous Hands
• Hardware: 1x RTX 4090

Problem Summary

I'm attempting to adapt VideoMimic for a G1 robot equipped with Inspire Dexterous hands, but encountering issues both with the pretrained model and retraining from scratch.

Issue 1: Pretrained Sim2Real Model Unstable with Dexterous Hands

Observation: When loading the pretrained sim2real policy on the modified URDF (G1 + Inspire hands), the robot leans forward significantly and cannot maintain balance.
Using g1_with_inspire_hands.urdf with fixed wrist and hand joints.

Question: Are the pretrained models sensitive to mass distribution changes? Should the models be retrained even for minor changes like adding hands?

Issue 2: Stage 1 Retraining Stagnates at Low Performance

Training attempts (both failed similarly):

Attempt 1: Original train_stage_1.sh script
Used default training script with modified URDF and only walking clips (lafan_replay_data/*walk*.pkl)

Attempt 2: Modified config using g1_deepmimic_mocap task and paper's rewards

• Task: g1_deepmimic_mocap (noticed this task exists but isn't used in original scripts)
• Config changes based on paper recommendations:

REWARD_ACTION_RATE = -8.0
REW_POS_ERROR_THRESHOLD = 0.3 
CONTACT_NO_VEL = -100.0
ALIVE_REWARD = 300.0
USE_ALT_FILES = False 
AMASS_TERRAIN_DIFFICULTY = 1 
...

• Walking clips only
• Added hand links to penalize_contacts_on

Both attempts show identical stagnation:

• Success rate: Plateaus at ~0.02 (2%)
• Episode length: Stuck around 30 steps
• Training duration: ~50,000 iterations over 3 days each

Questions for Authors

1. URDF Compatibility: Should Stage 1 policies be retrained when modifying the robot URDF (adding hands)? Or should the pretrained model transfer better?
2. Task Selection: I notice g1_deepmimic_mocap task exists but isn't used in train_stage_1.sh. What's the intended use case for this task? Should it be used for retraining with modified morphologies?
3. What should I do to resolve this stagnation? Any guidance on debugging approach, configuration changes, or expected training behavior?