Franka Research 3×openvla-7b-oft-finetuned-libero-spatial

Evidence

Success rate

82.0%

Episodes tested

500

Updated

2026-04-20

https://arxiv.org/abs/2502.19645

Notes

OpenVLA-7B + Optimized Fine-Tuning (OFT) on LIBERO-Spatial. Beats vanilla OpenVLA on this benchmark by a few points and is faster at inference because OFT prunes the action head.

Gaps noted

• · OFT (Optimized Fine-Tuning) fine-tune is LIBERO-Spatial-specific; cross-domain transfer to pour / fold tasks not measured in the cited paper.

Robot · Franka Research 3

Franka Research 3, 7-DoF torque-controlled research arm. Successor to Franka Emika Panda. De facto manipulation-research standard.

View robot profile →