TL;DR

When a robot exposes more joints than the policy was trained on (e.g. aloha has 16 joints but ACT expects 14), the policy raised a hard ValueError during inference. This fix auto-adapts state dimensions — truncate excess or zero-pad if fewer — with debug logging.

What changed

Why

This is the standard approach in robotics — LeRobot's own teleoperation code does the same. Hard crashing on dimension mismatch makes sim↔real transfer fragile and prevents running policies trained on one embodiment on another with different joint counts.

# Before: Hard crash
if len(state_values) != expected_dim:
    raise ValueError(f"State dimension mismatch: got {len(state_values)}...")

# After: Auto-adapt with logging
if len(state_values) > expected_dim:
    logger.debug("State dim %d > model expects %d — truncating", ...)
    state_values = state_values[:expected_dim]
elif len(state_values) < expected_dim:
    logger.debug("State dim %d < model expects %d — zero-padding", ...)
    state_values.extend([0.0] * (expected_dim - len(state_values)))

Testing

• ✅ All 266 existing tests pass
• ✅ Updated test_state_padded_to_expected_dim to verify auto-pad behavior
• ✅ Lint clean (ruff check + ruff format --check)

Part 1 of 6 in the MuJoCo simulation PR decomposition (see PR_TASKS.md)