place_empty_cup — MPC vs pure-VLA, BOTH at max_steps 400 (apples-to-apples)
Same standard as bread/canpot v3-valid (poselib + object_center MPC, completion_frames=2, fling detection > 2.0 m/s with 15-tick grace, run-until- 50-valid with flung seeds dropped & re-drawn). The ONLY difference: here MPC AND pure-VLA both use max_steps=400, so it's a fully fair comparison on identical valid seeds.
Result (env_success / 50 valid)
| method (both 400 steps) | success | flung |
|---|---|---|
| MPC (object_center + poselib) | 33/50 (66%) | 2 |
| pure-VLA (passthrough) | 34/50 (68%) | 0 |
Tied (VLA +1). With equal time budget and identical seeds, MPC correction gives no benefit on cup — consistent with bread/canpot. For reference, at 200 steps cup was MPC 29/50, VLA 31/50; doubling the horizon lifted both ~+4 but kept them tied.
comparison/cmp_seed<seed>_VLA-<r>_MPC-<r>.mp4: LEFT=VLA, MIDDLE=MPC,
RIGHT=MPC constraints; labels green=SUCCESS / red otherwise.