Does OpenVLA-OFT run on Franka Research 3?
works with caveats
Capabilities
| Capability | State | Notes | Evidence |
|---|---|---|---|
| Train / fine-tune | works | Repo collects FR3 demonstrations, exports RLDS, and serves its own fine-tuned checkpoints (easy50000, hard80000) with unnorm_key franka3_banana_plate_all_success. | ev-ww249-readme, ev-ww249-runtime |
| Real-robot inference | works with caveats | ev-ww249-runtime, ev-ww249-notes | |
| Simulation inference | no verified evidence | — | |
| Teleop + record | no verified evidence | — |
Caveats
- Community-reported only (single third-party repo, ww249/Openvla-oft-for-Franka-Research-3); not documented by the OpenVLA-OFT authors or maintainer-tested.
- One task only (pick up the banana and place it into the plate), one third-person camera, Cartesian-delta actions streamed at 10 Hz through the author's own libfranka controller.
- The author reports full-task success only after capping end-effector speed and moving in small steps, and notes manually pausing near the table and above the plate before resuming the policy.
Evidence 3 sources; the cell shows the highest tier
- T3 · community reportedsource-codeverified 2026-10-25 (today)
Repo README (Chinese): a real Franka Research 3 data-collection repo producing single third-person-camera OpenVLA-OFT fine-tuning data; action chunk 16 x 7 at 10 Hz in the fr3_link0 frame.
“这是 Franka Research 3 真实机械臂数据采集仓库,用于采集单第三视角相机的 OpenVLA-OFT 微调数据。”
https://github.com/ww249/Openvla-oft-for-Franka-Research-3/blob/3584446bd7191dcb59eaf13bfa50ec1f1f51ff15/README.md - T3 · community reportedsource-codeverified 2026-10-25 (today)
robot_runtime README: OpenVLA-OFT Franka runtime queries a policy server and streams Cartesian velocity commands to the FR3 via libfranka (--execute), following the OFT deploy recipe used in training.
“The runtime follows the OpenVLA-OFT deploy recipe used during training:”
https://github.com/ww249/Openvla-oft-for-Franka-Research-3/blob/3584446bd7191dcb59eaf13bfa50ec1f1f51ff15/robot_runtime/README.md - T3 · community reportedsource-codeverified 2026-10-25 (today)
README test notes dated 2026.7.19 (Chinese): with an end-effector speed cap and small per-step moves the full task succeeded; author also manually stopped near the table/above the plate before re-running the VLA.
“4. 设置机械臂末端执行器执行速度上限,每次移动小步,实现完整任务的成功”
https://github.com/ww249/Openvla-oft-for-Franka-Research-3/blob/3584446bd7191dcb59eaf13bfa50ec1f1f51ff15/README.md
Embed this cell
Badge for READMEs and model cards. It always shows the current state.
[](https://caniembody.pages.dev/compat/franka-fr3/openvla-oft)<iframe src="https://caniembody.pages.dev/embed/franka-fr3/openvla-oft" width="380" height="132" style="border:0" title="OpenVLA-OFT on Franka Research 3 | CanIEmbody"></iframe>History
- 2026-10-25 fecccd1 Initial release: robot policy compatibility matrix · Danilo Znamerovszkij