These are the weights of our final policy. It took 2nd place in the CoRL 2026 RoboEval Diagnostic Bimanual Challenge (Track 2), with a RoboEval Diagnostic Score (RDS) of 0.982 on the organizers' 1,000 held-out seeds per task.
The code, the method and the instructions to run it are on GitHub: github.com/Metric-AI-Lab/roboeval-bimanual-2026.
checkpoint/ holds the same files we entered in the challenge:
| File | What |
|---|---|
infer.json |
per-task settings: which model or pose network + scripted controller each task uses |
pose_*.pt |
7 pose networks (ResNet-18 image encoder + MLP), one per task |
policy.pt, policy_1.pt ... policy_8.pt |
fully learned policies (general and per task); only LiftTray's acts in the final policy |
To download them into a clone of the GitHub repo, run bash scripts/download_weights.sh.
checkpoint/policy_1.pt 114 MBcheckpoint/policy_2.pt 114 MBcheckpoint/policy_3.pt 114 MBcheckpoint/policy_5.pt 114 MBcheckpoint/policy_6.pt 114 MBcheckpoint/policy_7.pt 114 MBcheckpoint/policy_8.pt 114 MBcheckpoint/policy_4.pt 114 MBcheckpoint/policy.pt 114 MBcheckpoint/pose_shelf.pt 49 MBcheckpoint/pose_pot.pt 49 MBcheckpoint/pose_blocks.pt 49 MBcheckpoint/pose_box.pt 49 MBcheckpoint/pose_handover.pt 49 MBcheckpoint/pose_pick.pt 49 MBcheckpoint/pose_valve.pt 49 MB