China's BGI Subsidiary and Shanghai AI Lab Release ProtoPilot, First AI System to Complete Wet-Lab Experiments End-to-End, Outperforming GPT-5.6 Sol

While Silicon Valley AI giants are still having models write experimental protocols, a cross-industry player from Shenzhen has already put AI to work inside the actual lab.

On July 3, Yongsheng Intelligence, a subsidiary of BGI Group, together with the Shanghai Artificial Intelligence Laboratory, jointly announced two new achievements: ProtoPilot, a self-evolving multi-agent system driven by real laboratory scenarios, and BioLab Bench, the first end-to-end Agent evaluation system in life sciences that spans from user requirements to executable device commands. This represents the world's first complete loop from natural language experimental intent to physical wet-lab execution.

Third-party benchmark data shows ProtoPilot has surpassed OpenAI's current flagship GPT-5.6 Sol in end-to-end life science agent capabilities. The result signals that in the AI for Bio race, a Chinese contender has been the first to bridge the gap between generating protocols and actually running experiments.

Current AI applications in life sciences mostly stay at the understanding and analysis level. OpenAI released GPT-Rosalind for drug discovery, Google launched Co-Scientist embedding multi-agent systems into scientific reasoning, and Anthropic debuted Claude Science Workbench for research workflows. Yet all these frontier models have been stuck at the laboratory door — capable of writing plans but not producing results.

According to the ProtoPilot paper, translating an experimental intent into actual wet-lab operations requires traversing five layers: scientific intent, protocol design, standard operating procedures, device code, and finally physical execution with feedback correction. Each layer involves different ambiguity challenges, from biological logic representation to device SDK instructions — an extremely complex chain.

Yongsheng's breakthrough lies in making AI no longer just a screen-bound assistant but an actual executor that controls the complete pipeline from experiment design to device operation. This echoes Jensen Huang's CES remark about the ChatGPT moment for Physical AI — except the first entity to deliver a Physical AI answer in life science labs is not an AI giant but a Chinese Bio company venturing into AI.

For the industry, this means AI's role in life sciences is poised to leap from assisted analysis to autonomous experimentation. The next questions are whether ProtoPilot can be rapidly deployed across different lab environments and whether BioLab Bench can become an industry-standard evaluation framework.