arrow-hf's picture
v3-valid complete: MPC vs pure-VLA (pure-VLA >= MPC both tasks) + all videos
f256245 verified
|
Raw
History Blame Contribute Delete
1.07 kB

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%).