RoboSyn ACT fine-tuned on official + mined data (negative result)
Official RoboSynChallenge/ACT_sim_<task> checkpoints fine-tuned 20k steps (lr 1e-5, batch 32) on the
official data plus mined episodes from shifted domains (moved cam_high, other light, randomized physics).
Worse than the official checkpoints with temporal ensembling on the same 15 scenes: item_assembly 11 -> 4,
manipulate_pipette 10 -> 8, sample_loading 0 -> 0. Kept for reference; the mined data target the real robot.
Trained on the RoboSynChallenge simulation data (CC BY-NC-SA 4.0); same licence. Evaluated with the
official scripts/eval_policy.py (random setting, expert-feasibility filter). Numbers are success rates:
"100 eps" = seed 20261011 full render; "15/30 eps" = seed 0 screens (lazy render), paired with the
other configs on the same scenes. Gripper decoding matters a lot (see "gripper" below): the env takes the
gripper value as a 0..0.05 m finger target, the data stores 0..1 with a ~25-step closing ramp.
Inference adapter: policy/lerobot_ckpt in https://github.com/kozhukovv/robosyn_manip_challenge (private).