OMNI-Q table-setting checkpoints (Intel Physical AI Online Challenge, 2026-09-16)
Training outputs of https://github.com/The-Hidden-Canopy/OMNI-Q_Weird_Stuff_Machine
(MIT). Fetch with bash scripts/fetch_checkpoints.sh from that repo.
| folder | what | license |
|---|---|---|
smolvla_so101_table/ |
SmolVLA action expert fine-tuned (8000 steps, batch 8) on 40 demonstration episodes recorded from the OMNI-Q dual SO-101 MuJoCo table-setting scene. Base: lerobot/smolvla_base (Hugging Face LeRobot). |
Apache-2.0, as the base model and LeRobot; attribution above. Derived weights; the base model's terms apply. |
omni_planner/ |
IDA Omni reference body (omni_planner_r1_final.pt, 1.8 GB) used as the plan-step reasoner with fenced decoding, plus its training receipt. Trained by The Hidden Canopy on plan-grammar traces. |
The Hidden Canopy LLC; gated access — request and it is auto-approved. Not for redistribution outside the gate without permission. |
How they are used (see SUBMISSION.md in the repo): the policy leads every
single-arm PICK/MOVE from three cameras + arm state + language; the governed
contact primitive completes the step; the reasoner proposes which object /
which arm next and the governed core validates and completes the plan.
Measured: VLA-first 10-seed harness 8/10 resolved (12/16 with seeded sampling);
OMNI-advised 10-seed montage 10/10 on the recorded seeds.
Environment variables the repo's scripts read:
OMNIQ_VLA_CHECKPOINT=<path>/smolvla_so101_table
OMNIQ_OMNI_REASONER=omni OMNIQ_OMNI_CHECKPOINT=<path>/omni_planner/omni_planner_r1_final.pt OMNIQ_OMNI_RECEIPT=<path>/omni_planner/omni_planner_r1_final_receipt.json