RoboSyn StableVLA: per-task SmolVLA (smolvla_base, 50k steps) for each of the 10 tasks
Folders <task>_50k/. LeRobot smolvla policy, frozen VLM, chunk 50, 10 flow steps.
Raw gripper, 100 eps (seed 0): click 79, handle 61, water 69, table 90, handover 15, drawer ~6, mixer ~62,
assembly ~43, pipette ~38. Gripper decoding: with rescale drawer goes 1/30 -> 12/30 (hysteresis keeps the
fingers open until the command drops below 0.05, so the policy, which reads its own finger position,
restarts the closing ramp). With rescale_latch (15 eps): drawer 6, handle 8, handover 1, water 9.
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).