Does OpenVLA-OFT run on WidowX 250 S?

Capabilities

Caveats

• The WidowX result is an appendix experiment: OpenVLA-OFT (without FiLM) trained on all of BridgeData V2 and evaluated on a subset of OpenVLA's Bridge tasks (69.2% vs. 65.8% for OpenVLA).
• The openvla-oft repo ships fine-tuning and evaluation code only for LIBERO and ALOHA. It has no Bridge/WidowX config, deployment script or released WidowX checkpoint.

Evidence 3 sources; the cell shows the highest tier

1. T1 · upstream documentedpapersource 2025-04-28verified 2026-10-24 (1d ago)

   OpenVLA-OFT paper, Appendix (Scaling up to BridgeData V2): OFT is trained on BridgeData V2 and evaluated on real WidowX tasks from OpenVLA's evaluation suite, scoring 69.2% on average vs. 65.8% for OpenVLA.

   “Specifically, we train OpenVLA-OFT (without FiLM) on the BridgeData V2 dataset [52], which contains 50,365 real WidowX robot demonstrations (25× more than the four LIBERO task suites combined).”

   https://arxiv.org/abs/2502.19645v2
2. T1 · upstream documentedofficial-docssource 2025-09-09verified 2026-10-24 (1d ago)

   openvla-oft README points to LIBERO and ALOHA for fine-tuning/evaluation. experiments/robot at this commit contains only aloha/ and libero/.

   “See [ALOHA.md](ALOHA.md) for fine-tuning/evaluating on real-world ALOHA robot tasks.”

   https://github.com/moojink/openvla-oft/blob/e4287e94541f459edc4feabc4e181f537cd569a8/README.md
3. T1 · upstream documentedpapersource 2024-01-17verified 2026-10-24 (1d ago)

   BridgeData V2 was collected on a WidowX 250.

   “showing the WidowX 250 robot arm and various cameras.”

   https://arxiv.org/abs/2308.12952v3