What Is SimpleVLA-RL?

SimpleVLA-RL is introduced in SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning (arXiv 2509.09674) by Haozhan Li, Yuxin Zuo, Jiale Yu, Yuhao Zhang, Zhaohui Yang and co-authors, submitted on 11 September 2025. The abstract starts from two problems with supervised fine-tuning (SFT) of VLA models: large-scale human-operated robot trajectories are scarce and costly, and generalization under distribution shift is limited. It asks whether reinforcement learning, which improved step-by-step reasoning in large reasoning models, can similarly improve long-horizon action planning in VLAs.

SimpleVLA-RL is third-party research. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data. What the site states is custom egocentric collection of real human demonstrations, fast and on demand, to a written spec: 3D hand pose at 21 joints per hand, calibrated stereo depth and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. The only language data the site states is one task instruction per episode. Whether Firsthand data suits the SFT starting point or the RL stage of a SimpleVLA-RL-style pipeline is Not stated in the sources read.

Per the abstract, VLA models have progressed through large-scale pretraining and SFT, but SFT scaling needs large-scale human-operated robot trajectories that are scarce and expensive to collect. The models also generalize poorly to tasks with distribution shift. SimpleVLA-RL targets both by adding an RL stage after SFT.

The authors build on veRL and add VLA-specific trajectory sampling, scalable parallelization, multi-environment rendering and optimized loss computation. The README describes binary 0/1 outcome rewards, so no complex reward design is needed, plus exploration strategies: dynamic sampling, adaptive clipping and temperature tuning.

The README also states that an SFT VLA model is required before RL training. RL here refines a model that already works somewhat; it does not replace the SFT stage.

Applied to OpenVLA-OFT, the abstract reports state-of-the-art performance on LIBERO, better results than π0 on RoboTwin 1.0 and 2.0, and gains over SFT on real-world tasks. The authors also describe a “pushcut” phenomenon, where the policy discovers patterns not seen in earlier training.

These are the authors’ results on their own setups. Firsthand does not reproduce them and quotes no scores.

The official repository, PRIME-RL/SimpleVLA-RL, is released under the MIT License. The README lists OpenVLA and OpenVLA-OFT as supported VLA models and LIBERO and RoboTwin 1.0/2.0 as benchmarks, with SFT checkpoints available from its Hugging Face collection.

The site’s OpenVLA-OFT entry covers a supervised fine-tuning recipe, and its LoRA entry covers parameter-efficient adaptation. SimpleVLA-RL is a later stage on top of an SFT model, trained with reward signals from environment rollouts. All are third-party methods, and Firsthand ranks none of them.

This entry draws on the arXiv abstract page and the official GitHub repository. Neither describes human egocentric video, stereo depth, hand pose or IMU data in SimpleVLA-RL’s evaluation, so those are Not stated.

When a program needs real human egocentric demonstrations that public data does not cover, specified in writing. Firsthand does not build or train models. Confirm the spec on a Spec Pilot, a small fast first batch; the median from signed spec to first validated batch is 48 hours where coverage is deep, three to four weeks where thin.

What is SimpleVLA-RL?

A published RL framework for VLA models (arXiv 2509.09674, ICLR 2026), built on veRL, that fine-tunes an SFT VLA model such as OpenVLA-OFT with binary outcome rewards.

Does SimpleVLA-RL replace supervised fine-tuning?

No. The README states an SFT VLA model is required for RL training, so RL is applied after SFT.

What did the authors report?

On OpenVLA-OFT, the abstract reports state-of-the-art results on LIBERO, better results than π0 on RoboTwin 1.0 and 2.0, and gains over SFT on real-world tasks. These are the authors’ results.

Does Firsthand data train SimpleVLA-RL?

No claim is made. Firsthand is not affiliated with its authors, and the site does not state that any model was trained on Firsthand data.

Can Firsthand collect egocentric data for a VLA program?

The site states custom egocentric collection to a written spec, fast and on demand, with 21-joint 3D hand pose, calibrated stereo depth, 200 Hz IMUs and one task instruction per episode, confirmed on a Spec Pilot.