PHC

Project notes

Choose the policy before copying an evaluation command

PHC's single-primitive model can track motion without the full mixture, but upstream explicitly says it lacks failure-state recovery. Upstream describes the PHC+ keypoint configuration as able to get up and walk back; its rotation-plus-keypoint configuration can get up but not walk back. Keep the checkpoint, robot, observation settings and environment from one documented command together. [1]

For a first evaluation that needs both getting up and walking back, our suggested starting point is the README’s PHC+ keypoint command: exp_name=phc_comp_kp_2, the primitive checkpoint output/HumanoidIm/phc_kp_2/Humanoid.pth, robot=smpl_humanoid, env.obs_v=7 and env.num_prim=3. Its composer checkpoint is downloaded under output/HumanoidIm/phc_comp_kp_2/. Keep the complete command’s other settings, including the standing sample sample_data/amass_isaac_standing_upright_slim.pkl, env.num_envs=1 and headless=False. First prepare the documented Python 3.8, PyTorch/CUDA 11.6 and Isaac Gym environment, then obtain the body models and matching checkpoints. Inspect download_data.sh before using it: it fetches multiple policies and samples, rather than only the files needed for one example. These are upstream setup instructions, not a reproduced installation. [1] [3]

Coordinate conventions affect the result

The released models use a modified SMPL convention: gravity points along negative z and the humanoid faces positive x. The README calls out up_right_start and the conversion work needed when moving back to SMPL. Check this before interpreting a failed imitation as a policy defect. [1]

The reported AMASS results use a particular body shape, height adjustment and evaluation protocol. Upstream notes that batch size and machine setup can change results. We have not reproduced those success rates, and the inspected setup does not establish a minimum GPU-memory budget for a chosen workload. [1]

Check external artifacts separately

This profile pins commit 846988d433ce1f341e85ac6fbd2cd51911bb3341; no GitHub release was listed. Several dependencies follow Git branches, so the PHC commit alone is insufficient to reproduce the environment. [2] [6]

The source license is BSD-3-Clause-Clear and points to separate asset licenses. SMPL models, AMASS motions and externally hosted checkpoints are separate artifacts whose permissions are not established by the root code license. The download script documents those locations; it is not a license grant for them. [7] [1] [3]

We inspected source and made limited artifact-access checks. A sample-data request returned a partial binary response, while a checkpoint request reached a Google Drive download-confirmation page. Complete payloads, checksums and model loading were not verified. No training, simulator evaluation or hardware test was performed. [8]