ModelFeatured
Octo
Generalist robot policy Transformer with multi-modal inputs and multi-embodiment finetuning.
93MFully openData: Open X-EmbodimentUpdated 2024-07-31
ProjectFeaturedIntermediate
OpenVLA-OFT
Optimized Fine-Tuning (OFT) for VLAs (RSS 2025). Boosts OpenVLA on LIBERO (~76.5%→97.1%) with parallel decoding, action chunking, and continuous actions; strong real ALOHA results.
Why: 2025 OpenVLA follow-up: OFT finetune recipe + paper arXiv:2502.19645
Overview
OpenVLA-OFT is the 2025 follow-up to the base OpenVLA model paper. Paper: Fine-Tuning Vision-Language-Action Models (arXiv:2502.19645, RSS 2025). It studies action decoding, continuous actions, and learning objectives, and proposes an Optimized Fine-Tuning (OFT) recipe.
Reported gains include lifting OpenVLA’s LIBERO average success from about 76.5% to 97.1% while increasing action throughput substantially, plus strong real bimanual ALOHA results versus π0 / RDT-style finetunes. If you already use OpenVLA weights, this is usually the first paper/code to read after the 2024 base paper.
Who it is for
Teams finetuning OpenVLA on limited demos; readers wanting the 2025 OpenVLA-line paper.
Key highlights
- 2025 RSS paper arXiv:2502.19645
- LIBERO SOTA-style gains for OpenVLA finetune
- Parallel decoding + action chunking + continuous actions
- Open code and checkpoints
When to use
- Finetuning OpenVLA on small-to-medium robot datasets
When not to use
- You are training a non-OpenVLA architecture from scratch
Getting started
- 1Start from OpenVLA checkpoint
- 2Follow OFT configs
- 3Ablate learning rates on a tiny split
Papers & reading
How it compares
ProjectFeatured
openpi
Open training and inference code around π0 from Physical Intelligence for generalist robot policies.
Caveats & pitfalls
- Still inherits OpenVLA license and compute needs.
Content reviewed 2026-07-23