OpenVLA-OFT
The practical fine-tuning recipe for OpenVLA: LoRA plus continuous action heads that adapt the 7B VLA to new robots fast.
About
OpenVLA-OFT (Optimized Fine-Tuning), from Stanford researchers in February 2025, is the recipe that made OpenVLA actually usable for robotics labs. Instead of full fine-tuning, it adapts the 7B vision-language-action model with rank-32 LoRA, replaces tokenized actions with continuous action heads (L1 regression or diffusion), and adds action chunking, proprioceptive inputs, and optional FiLM conditioning for stronger language grounding. The OFT+ variant takes three camera views for real-world setups like ALOHA.
The authors reported 95–98% success across the LIBERO simulation suites (Spatial 97.2%, Object 97.4%, Goal 95.8%, long-horizon LIBERO-10 98.0%) with far less compute and data than prior VLA fine-tuning approaches. Community projects have since used the recipe to put OpenVLA policies on UR5e arms and other hardware through ROS2 pipelines. It's a research artifact — code and checkpoints, not a product — and the exact license terms of the derived checkpoints deserve a check before commercial use, but for VLA fine-tuning it's the reference method.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- LIBERO-Spatial 97.2%, LIBERO-Object 97.4%, LIBERO-Goal 95.8%, LIBERO-10 98.0% (author-reported, OpenVLA-OFT paper setup, arXiv 2502.19645)
- No independent benchmark replication verified
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: peer paper arXiv 2502.19645 (Feb 2025) by Kim/Finn/Liang; open reference implementations on GitHub; community projects (UR5e pick-and-place via ROS2) corroborate real-world use; LIBERO numbers are author-reported in the paper setup. Open weights/checkpoints. Strengths: the standard method for adapting OpenVLA, dramatically cheaper than full fine-tuning, proven on real robots by third parties. Docked for: (1) benchmark numbers are author-reported with no independent replication found; (2) license terms of derived checkpoints not clearly documented — verify before commercial use. License: not verified — check the repo before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Paper arXiv 2502.19645 (Feb 2025) by Moo Jin Kim, Chelsea Finn, Percy Liang (Stanford)
- · LoRA rank-32 + continuous action heads (L1/diffusion) + action chunking + proprioception + FiLM
- · Author-reported LIBERO: Spatial 97.2%, Object 97.4%, Goal 95.8%, LIBERO-10 98.0%
- · Reference fine-tuning method; used in community real-robot projects (UR5e + ROS2)