BCIwiki (bciwiki.com) — A framework called BrainVLA decodes motor intention by drawing on a pretrained vision-language-action (VLA) model, outperforming the evaluated baselines in cross-session decoding R² and task success rate on 2 neural motor datasets, according to a preprint posted to arXiv on September 28, 2026. The authors are Luyao Jin, Running Zhao, Huan Zhao, Vincent C. K. Cheung and Wei-Hsin Liao. The work has not been peer reviewed.
Motor brain-computer interfaces (BCIs) aim to let people with paralysis control external devices. Neural motor decoding typically learns task-specific mappings from neural activity to kinematics, an approach constrained by scarce paired neural-action data. BrainVLA seeks to reduce that dependence by leveraging VLA policies.
The researchers first built VLA-compatible datasets pairing neural activity with action signals, language instructions and rendered visual observations. They then adapted the OpenVLA-OFT policy to the target action spaces through LoRA fine-tuning. Finally, they trained a neural encoder via neural-language alignment, using language representations as semantic targets to capture latent motor intent. The resulting representations act as an endogenous intention signal guiding the VLA policy to generate executable actions, while visual observations supply information about the evolving task state.
Evaluation used causal rollout decoding on 2 datasets with different action dimensionalities. Besides beating baselines, BrainVLA showed high training data efficiency, the abstract says. The authors say the results establish a route for neural motor decoding to draw on large-scale robotic priors. The abstract gives no specific figures and does not describe how the datasets were collected.