v3-valid eval — fling-excluded, MPC vs pure-VLA (latest code 815ea9c)
Latest vla_ppc (bbox-centre-min distance + completion_frames=2). 50 valid (non-flung) episodes per task for MPC; flung episodes (task object ejected by a physics glitch, GT speed > 2.0 m/s) are dropped and shown separately. pure-VLA ran on those same MPC-valid seeds. SmolVLA, max_steps 200, chunk_exec 50.
Results (env_success)
| task | MPC v3-valid | pure-VLA (same seeds) |
|---|---|---|
| place_bread_skillet | 25/50 (50%) | 25/46 (54%, 4 flung) |
| move_can_pot | 38/50 (76%) | 41/49 (84%, 1 flung) |
pure-VLA ≥ MPC on both tasks — MPC correction does not help (slightly hurts) these tasks; SmolVLA alone is best. The latest distance/de-bounce code did not change this.
Layout per task: mpc/ (poselib + MPC: bread=economicgrasp, canpot=object_center)
and pure_vla/ (passthrough, no MPC). Each has videos/ (counted episodes) and
flung_videos/ (dropped; 15-tick grace so the fling is visible). bread flings
often (34% of seeds); canpot almost never (2%).