Policy weights for our RoboSynChallenge 2026 real-robot finals submission (team OC-Manip)
Thank you to the RoboSynChallenge organizers for the benchmark, the simulator, the data and the real-robot evaluation.
This repository holds the weights used by our finals policy. The code is in
NeurIPS2026-RoboSynChallenge-Final-Round-Submission. The code pins this repository at revision 2a395d9680be8466a0a554a8fb888b2a82aada1b;
please use that revision.
Layout
shared/vjepa2_1_vitl_384_ema_encoder.pt frozen V-JEPA 2.1 ViT-L/16-384 encoder, used by every task
<task>/README.md short description of the task's files
<task>/student/ the task's policy (OfI, One from Infinity)
ofi_checkpoint.json file list with sha256 (checked before loading)
step_<N>.pt policy weights
config.yaml, dataset_stats.json training configuration, normalization statistics
meta/tasks.jsonl, text_cache/*.pt task instruction and its embedding
<task>/helper/*.pt keypoint heads of a task's helper (sample_loading, click_bell)
Tasks
Success counts are from our own evaluation in the official simulator with the wrist cameras at the real-robot mounting angle; the real robot was not tested.
| Task | Policy | Training step | Weights sha256 (first 8) | Simulator success / episodes | Helper files |
|---|---|---|---|---|---|
| Click Bell | click_bell/student/ |
160000 | 02a2ed93 |
7 / 20 | click_bell/helper/head.pt, click_bell/helper/tip_head.pt (off) |
| Drawer Open Place | drawer_open_place/student/ |
160000 | 91619ff7 |
2 / 34 | none |
| Handle Basket | handle_basket/student/ |
160000 | 6c6f1e52 |
0 / 25 | none |
| Item Assembly | item_assembly/student/ |
160000 | 1db9a74c |
4 / 20 | none |
| Items Handover | items_handover/student/ |
160000 | ca4b805f |
2 / 28 | none |
| Manipulate Pipette | manipulate_pipette/student/ |
120000 | 0222dd54 |
6 / 36 | none |
| Mixer Operating | mixer_operating/student/ |
180000 | 733f5d97 |
3 / 43 | none |
| Sample Loading | sample_loading/student/ |
160000 | 917d9285 |
3 / 49 | sample_loading/helper/head.pt (on) |
| Table Rearrangement | table_rearrangement/student/ |
160000 | ae7a3079 |
46 / 50 | none |
| Water Pouring | water_pouring/student/ |
160000 | c7b98d4a |
12 / 39 | none |
The encoder shared/vjepa2_1_vitl_384_ema_encoder.pt has sha256 49231ae6951961715ccdf993aadd755b8d5f27864e130b1129bef0305e0edcf0.
Use
Follow the setup in the code repository. Weights are gated; authenticated access is granted to the RoboSynChallenge account. One command downloads every pinned file into the local Hugging Face cache and checks each sha256:
cd RoboSynChallenge
policy/act/.venv/bin/python policy/MyPolicy/prefetch.py
It ends with [prefetch] done: N of N and exits with an error if anything is missing.
Acknowledgements and licenses
- The policy model is OfI from RoboActualizer (MIT).
- The shared encoder is the V-JEPA 2.1 ViT-L/16-384 checkpoint from
V-JEPA 2 (MIT; see
shared/LICENSE-vjepa2). - The instruction embeddings in
text_cache/were computed with the Wan2.2 text encoder (umT5, Apache-2.0).
Contact
Yi Li (team OC-Manip). Please open an issue in the code repository.