A research team has built a multi-agent robotic system that automates biological lab procedures by reading standard protocols and verifying each step before moving on.
BioProVLA-Agent breaks the job into three cooperating components: one agent parses a written protocol into discrete, checkable subtasks; a second uses a vision-language model and a retrieval system to confirm each step succeeded before the next begins; and a third executes the physical movements via a lightweight action policy. The team also developed AugSmolVLA, an augmentation strategy that trains the vision system to handle the specific hell of wet-lab optics — transparent tubes, reflective surfaces, blown-out lighting. They tested the system on a benchmark of 15 atomic tasks, 6 composite workflows, and 3 bimanual tasks, including tube loading, sorting, waste disposal, cap twisting, and liquid pouring. Against comparable baselines — ACT, X-VLA, and the original SmolVLA — their approach held up better under normal and high-exposure conditions alike.
Lab automation has long been a domain for expensive, purpose-built machines locked to specific workflows. A system that accepts plain protocols as input and closes the loop with visual verification could let smaller labs automate without buying a $200,000 liquid-handling robot. The closed-loop design also addresses a real failure mode: robots that barrel through steps regardless of whether the last one worked.
The paper is careful to say this "suggests a practical route" rather than declaring the problem solved — which is about the right level of confidence for a benchmark that does not yet include the chaos of a real shared lab bench.