| # Side-by-side: pure-VLA vs MPC (MPC updated with 400-step re-runs) | |
| Filename + in-video labels encode both outcomes (green=SUCCESS, red otherwise): | |
| `cmp_seed<seed>_VLA-<result>_MPC-<result>.mp4`. LEFT = pure VLA (200 steps), | |
| MIDDLE = MPC, RIGHT = MPC constraint readout. | |
| **MPC results were refreshed**: seeds that MPC failed at 200 steps were re-run at | |
| 400 steps; the recovered ones now show SUCCESS and use the 400-step MPC video. | |
| So MPC here is "up to 400 steps", VLA is 200 steps (not fully apples-to-apples). | |
| ## Final success (env_success / 50 valid seeds) | |
| | task | MPC (≤400 steps) | pure-VLA (200 steps) | | |
| |---|---|---| | |
| | place_bread_skillet | 33/50 (66%) | 25/50 (50%) | | |
| | move_can_pot | 39/50 (78%) | 41/50 (82%) | | |
| Updated MPC breakdown: bread 33 SUCCESS / 7 FAIL / 10 FLUNG; canpot 39 SUCCESS / | |
| 11 FAIL / 0 FLUNG (the 400-step horizon made bread fling far more). | |