[Alpenj]

Summary

• add --parallel-envs for one batched SmolVLA policy driving asynchronous MuJoCo workers
• isolate per-worker RCS-compatible traces and resolve evaluator episodes to traces by reset seed
• emit a machine-readable rollout manifest that distinguishes experience generation from optimizer training
• default unattended MuJoCo subprocesses to EGL and document measured behavior

Measured result

Using seeds 1800–1803 and a 700-step horizon on the RTX 5050 Laptop GPU 8 GB:

• sequential 1-env: 4/4 success, 30.726 s, 1,505 trace frames
• parallel 4-env: 3/4 success, 31.763 s, 2,800 trace frames

This is not claimed as a speedup or training update. It is useful for bounded failure harvesting; reproducible evaluation remains batch size 1. Optimizer updates in this stage: 0.

Verification

• 40 passed, 1 warning for the SO101 MuJoCo test suite
• Ruff checks passed
• EGL 2-worker/5-step checkpoint smoke passed without the earlier X11 cleanup error
• 2 workers × 2 batches smoke verified seeds 1820–1823 map to the correct worker/local trace episodes