Determinism and versioning

Bit-exact replay

Restoring a saved simulator state mid-episode reproduces the rest of the trajectory to the last bit.

• The G1 scenes use the Newton solver. PGS warmstarts from contact-dependent state that cannot be saved.

• A snapshot holds the full controller state: commands, last action, gait phase and the policy’s observation history.

• The env runs mj_forward after every control step and after a restore, so kinematics, contacts and sensors match.

Each demonstration stores its engage snapshot: init_qpos, init_qvel, init_ctrl, init_qacc_warmstart and the controller state (lb_state.*). Replay resets the env on the episode’s seed, restores the snapshot and applies the recorded actions.

Deterministic reset

The official configuration sets deterministic_reset, so every reset restores the controller state from the env’s first reset and reset(seed) depends only on the seed.

Substrate version and fingerprint

SUBSTRATE_VERSION is bigym2-mj381-v1. It changes when the MuJoCo version, the physics options, the observation and action layout, or the reward and success semantics change.

Official results are comparable when their substrate_version matches. Official means protocol_violations(env) is empty (see Official overrides).

env.substrate_fingerprint() adds the identity of the exact run: hashes of the compiled model and the lower-body weights, the task and controller settings, bigym_version and the git SHA. Leaderboard records, demo batches and coding-agent evaluations embed it.

The fingerprint does not cover Python packages. Official results use Python 3.12 with the versions pinned in uv.lock.