MGI Tech and Shanghai AI Lab Launch Physical AI for Life Sciences

The newly introduced ProtoPilot acts as a self-evolving multi-agent system capable of managing the entire experimental lifecycle, from design to wet-lab feedback. By learning from real-world failures, such as diagnostic errors in antibiotic screening, the system autonomously regenerates corrected protocols. Performance metrics underscore its potential; on the ProtocolQA benchmark, ProtoPilot reached 52.38% accuracy, nearing the 54% threshold typically held by human experts.

Complementing this is BioLab Bench, an evaluation framework that moves beyond theoretical accuracy to assess whether an AI can successfully operate physical automation hardware. By testing agents across three difficulty levels, the bench evaluates the full chain of execution—from intent interpretation to device-specific machine code production. This initiative, spearheaded by Genoria AI CEO Dr. Yang Meng, shifts the focus from pure computational scale to a closed-loop engineering model. By integrating expert validation and hardware-native data, the collaboration aims to standardize the development of 7×24 unattended intelligent laboratories.