Request for Guidance on Humanoid Evaluation Instability

[cold-deuu]

Hi ZhengyiLuo,

First, thanks for releasing the code. I trained and evaluated H1 with the following commands:

the train command:

python phc/run_hydra.py \
  project_name=Robot_IM \
  robot=unitree_h1 \
  env=env_im_h1_phc env.motion_file=sample_data/sample_dance_h1.pkl \
  learning=im_pnn_big \
  exp_name=unitree_h1_pnn_realsim_092924 \
  sim=robot_sim control=robot_control \
  learning.params.network.space.continuous.sigma_init.val=-1.7

and evaluation command

python phc/run_hydra.py \
  learning=im_pnn_big exp_name=unitree_h1_pnn_eun \
  epoch=-1 test=True \
  env=env_im_h1_phc robot=unitree_h1 \
  robot.freeze_hand=True robot.box_body=False \
  env.motion_file=sample_data/dance_sample_h1.pkl \
  env.num_envs=1 headless=False \
  control=robot_control

During evaluation/visualization, the robot sometimes gets “catapulted” off the ground as if it receives a strong contact impulse, and control breaks down.

I’m mainly wondering about two things:

1. Will this disappear with more training? I trained on a single A6000 GPU for 3 days. Is this instability something that typically resolves as the policy trains longer, or should I expect stability already at this stage?
2. Could this be a simulation setup issue? For example, ground/foot restitution, friction, max depenetration velocity, bounce threshold, timestep/substeps, or solver iterations causing large initial contact impulses.

If it’s neither of the above, I’d greatly appreciate any additional pointers you can share (e.g., a known-stable eval config for H1, recommended PD gains/action scaling at eval, foot geometry/material settings, or any differences between your training vs. eval physics settings that are important).

I can also share a short video if helpful. Thank you! Best regards, Seulchan Lee

https://github.com/user-attachments/assets/401f8d3f-4bf9-4300-bf0c-2367cd8117ec