Fix floating-base fail check replay divergence

[ZidongChen25]

Summary

Fix a replay divergence in floating-base environments where the same action sequence could succeed with env.step(action, fast=True) but fail with env.step(action, fast=False).

Root cause

BiGymEnv._fail() previously checked whether the robot moved too far from the origin by reading the pelvis Cartesian position through self._robot.pelvis.get_position(). For the pelvis body this accesses MuJoCo's derived xpos field through dm_control.

When the physics state is dirty, reading a derived MuJoCo field can trigger an implicit physics.forward(). That extra forward pass does not advance simulation time, but it can recompute derived state, contacts, and constraint data. In contact-heavy tasks such as StackBlocks, doing this during every fast=False step can perturb the later physical trajectory.

Change

For floating-base robots, _fail() now reconstructs the pelvis XYZ position from the floating-base joint qpos values instead of reading body xpos. For axes that are not actuated, it falls back to the static MJCF body pos. Non-floating robots keep the previous pelvis.get_position() behavior.

This preserves the original fail check semantics while avoiding a derived-field read that can force an extra forward pass.

Validation

• python -m py_compile bigym/bigym_env.py
• Validated on a StackBlocks replay that previously diverged: after this change, the fast=False replay reaches success at frame 1344, matching the successful replay path.

[stepjam]

LGTM! Thank you!