ProtoPilot, a multi agent system that turns a scientist's intent into liquid handling robot commands, and BioLab Bench, the end to end benchmark that tests it, claim the missing link between language models and liquid handling robots, the same gap
The hardest step in putting AI inside a biology lab is not getting the model to design an experiment. It is getting a liquid-handling robot to actually run it. A team from Shenzhen and Shanghai says it has solved the last-mile problem that has kept even the most powerful language models confined to the computer screen.
That team is Yongsheng Intelligence (涌生智能), a subsidiary of the Chinese sequencing-instrument maker MGI Tech (华大智造), working with the Shanghai AI Lab (上海人工智能实验室). Together they have released two artifacts: ProtoPilot, which they describe as a self-evolving multi-agent system that takes a researcher's natural-language intent and walks it all the way to commands a lab robot can execute; and BioLab Bench, an evaluation suite designed to measure whether such an agent actually succeeds end-to-end. The original qbitai feature on the release frames the package as a "Physical AI" answer for life sciences, borrowing the framing NVIDIA chief Jensen Huang used at CES when he said the Physical AI ChatGPT moment had arrived.
The gap they are targeting is concrete, and it is the same gap that has tripped up Western AI-for-bio efforts. In the OpenAI partnership with Ginkgo Bioworks, for example, GPT-5 handles experimental design and parameter exploration. The Catalyst protocols that actually run on the bench, however, are still authored by Ginkgo's human engineers. AI proposes; humans translate. The qbitai write-up and its Sina Finance pickup describe that split as a "design-not-execute" ceiling the new system is meant to clear.
ProtoPilot's stated ambition is to walk the full conversion chain: a scientist's stated intent becomes a written protocol, which becomes a standard operating procedure, which becomes the device-level code that pipettes liquids or sets incubation times, which then runs and feeds results back into the model. BioLab Bench is the harness for measuring whether each of those handoffs holds. The Shenzhen Securities Times summary describes the system as anchored to real biological experiment scenarios rather than to benchmark puzzles.
Several claims in the announcement deserve precision. The originating coverage in Shenzhen News calls the system the first life-science agent to close the dry-wet loop, language the vendor used but which no independent third-party test has yet confirmed. The same coverage teases a benchmark comparison to "OpenAI's flagship GPT-5.6 Sol" that the available reporting does not substantiate. Outside confirmation would be needed before that comparison should anchor any story. Jensen Huang's CES "Physical AI" line is best read as keynote framing the announcement borrowed, not as endorsement of this specific system.
It also helps to keep the corporate geography straight. MGI Tech (华大智造) is the BGI Group's listed sequencing-and-instrument arm, distinct from BGI Genomics (华大基因), the sequencing-services parent. Yongsheng Intelligence is its Shenzhen-based AI subsidiary. The Shanghai AI Lab is a city-backed research institute, not a startup. The NBD strategy write-up frames this release as one step in MGI's broader AI-driven product roadmap, not as a standalone science result.
What would actually falsify or confirm the claim is narrow and specific: a peer-reviewed release of BioLab Bench with reproducible tasks; a side-by-side run against an existing AI-for-bio baseline on the same wet-lab hardware; and public access to the failure logs when an agent gets a step wrong. Until those land, the story to track is whether the conversion chain ProtoPilot promises, intent to protocol to device code to physical execution to feedback, can survive contact with a real bench. Whether it can outflank any particular competitor is a less interesting question.