HumanoidVerse

HumanoidVerse provides humanoid-policy training workflows across several simulators, with PPO implementation and configurable robot/task settings. Its source makes a concrete training framework available to inspect. It does not establish simulator equivalence, successful policy transfer or a commissioned physical humanoid. [1] [5]

The README deliberately separates environments: Python 3.8 for the IsaacGym route and Python 3.10 for IsaacSim/Genesis routes. It says readers need not install every simulator. The package metadata separately declares Python 3.8 or later and common dependencies including numpy 1.23.5 and Hydra. Those declarations do not erase simulator-specific version and GPU requirements. No environment was installed here. [1] [4]

For a bounded IsaacGym investigation, the README installs the separately supplied simulator Python API, then HumanoidVerse with pip install -e .. Its initial train_agent.py example selects simulator=isaacgym, locomotion, no domain randomization and one environment with a visible window. It matches +robot=h1/h1_10dof with +terrain=terrain_locomotion_plane, +obs=loco/leggedloco_obs_singlestep_withlinvel and +rewards=loco/reward_h1_locomotion; inspect that complete configuration before increasing the environment count or changing the robot. These commands were not run. [1]

The Genesis route names genesis-world 0.2.1 and notes that features such as recording are still unstable. The evaluation route takes a checkpoint and reloads saved configuration. The README advertises Sim2Sim/Sim2Real support and reports cross-simulator evaluations, while its roadmap still leaves those pipelines unchecked. Pipeline completeness and transfer behavior remain unresolved in this source-only review. A readable checkpoint filename or another simulator option is not proof of matched observations, dynamics, assets or transferable policy behavior. [1]

The inspected PPO excerpt contains model construction, optimizers, storage and a training loop calling _training_step; that excerpt does not show the called update method implementation. No training, rollout, optimization result or checkpoint evaluation was reproduced. MIT covers the project contribution; external simulators, robot assets, motion data and checkpoints require their own rights review. Begin with one simulator and the minimal configuration, then establish its dependencies and outputs before scaling or hardware integration. [2] [4] [5]