When giants like OpenAI and Anthropic direct their large language model capabilities towards life sciences, the competition over whether AI can truly take over the laboratory is quietly intensifying. Recently, the Shanghai AI Laboratory and MGI Tech Co., Ltd. (ASX: 688114) subsidiary Yongsheng Intelligence jointly released the ProtoPilot self-evolving multi-agent system and the BioLab Bench full-process evaluation framework, bringing the concept of "Physical AI for Life Sciences Laboratories" to reality for the first time. This signifies not only AI's progression from "reading papers" to "conducting experiments," but also marks a shift in lab automation from traditional mechanical automation towards "embodied intelligence" with perception, decision-making, execution, and self-evolution capabilities.
From Design to Execution: Physical AI Breaks the Agent Implementation Bottleneck
Over the past year, AI for Biology has become a global focus for tech investment and development. OpenAI launched GPT-Rosalind, Google released Co-Scientist and ERA, and Anthropic introduced Claude Science—applications of large models in life sciences have rapidly expanded from "knowledge layer" capabilities like literature reading and sequence analysis to downstream tasks such as hypothesis generation and experimental design.
However, a critical bottleneck persists: the ability to "design" an experiment does not equate to the ability to "perform" it. Even the most powerful models currently recognized by the industry remain confined to the "design layer" when faced with the physical execution of real-world laboratory work.
This is precisely the boundary that "Physical AI for Life Sciences Laboratories" aims to break. Unlike traditional AI agents confined to the digital realm, Physical AI emphasizes that intelligence must develop through interaction with the real world. Autonomous driving capabilities are honed on real roads, robotic skills are refined through real movements, and the competence of life sciences Physical AI must be cultivated within the real-world constraints of samples, reagents, consumables, plate positions, pipetting volumes, temperature conditions, equipment scheduling, and anomaly handling.
This is the core value of ProtoPilot. The system is not a single chatbot but is currently one of the few agent systems in the industry that comprehensively covers the entire chain from "experimental intent understanding → protocol generation → code conversion → equipment execution → wet-lab feedback verification." Researchers need only describe their experimental intent in natural language; the system can then decompose it into a scientifically sound experimental plan, translate it into executable SOPs and machine code, and continuously self-correct and evolve based on feedback from wet-lab results.
Wet-Lab Closed-Loop Validation: The 'Real World' Litmus Test for Bio Agents
Capital market enthusiasm for AI concepts often stops at "generating impressive answers," but the uniqueness of life sciences lies in the fact that any parameter error or equipment incompatibility can lead to experimental failure. Therefore, the moat for Physical AI lies not in parameter scale, but in the real-world capability of its "physical half."
As a life science tools company with over a decade of experience in the field, MGI's exploration of AI dates back to 2019. Last year, MGI's Yang Meng team, in collaboration with Professor Nattiya Hirankarn from Chulalongkorn University in Thailand, published a paper in Nature Biomedical Engineering. They developed a dry-wet collaborative multi-agent system named "PrimeGen," which innovatively integrated primer design, experimental validation, and automated workstation execution into a closed-loop process.
Furthermore, advantages such as native hardware compatibility, experience accumulated from serving over 3,800 global users, and engineering implementation expertise mean that MGI's AI models have been immersed in real experimental scenarios from their inception.
The joint team's validation of ProtoPilot did not stop at offline scoring; they actually executed the system-generated protocols in wet-lab experiments. More significant for the industry is the system's self-correction capability. When an anomaly occurred during the PCA assembly experiment conversion step, ProtoPilot could analyze the cause of failure, identify the ineffective resistance screening, and regenerate a corrected plan. This indicates the system has begun to possess the closed-loop capability of "learning from failure," a core differentiator of Physical AI from traditional lab automation software.
The Physical AI for life sciences laboratories developed through this collaboration points directly to the next paradigm of life science discovery. In the future, Bio Agents will no longer rely solely on text training to improve their capabilities. Instead, they will leverage the experimental chain established by Physical AI to continuously accumulate real-world research tasks, automated operations, expert review, failed samples, and on-site wet-lab feedback. After iterative optimization with vast amounts of physical experimental data, BioAgents will evolve to possess integrated reasoning, practical operation, and validation capabilities, enabling the implementation of 7x24 unattended intelligent laboratories.
MGI's 'Second Growth Curve': Fully Integrating AI Technology into Life Science Tools
For MGI, this joint release holds deep strategic significance. The market has long positioned the company as the "domestic alternative leader in gene sequencers." However, in recent years, the company has been accelerating its transition from this role towards becoming a "life sciences intelligent infrastructure platform." In April this year, MGI established a subsidiary focused on the AI for Science (AI4S) field—Shenzhen MGI Yongsheng Intelligent Technology Co., Ltd. (Yongsheng Intelligence)—dedicated to building a dry-wet closed-loop infrastructure for life sciences.
The released ProtoPilot and BioLab Bench are key moves in this strategy. The latter provides the industry with a unified benchmark for evaluating real experimental chains, and the party that sets the standard first gains a first-mover advantage in ecosystem building. The former, for the first time, demonstrates that Bio Agents can truly move from digital intelligence to the physical execution of experiments, closing the loop between experimental intent, protocol, code, and wet-lab feedback.
For MGI, its automation equipment matrix is no longer merely an execution terminal but has become an intelligent node capable of iterative learning. Every real experiment and its feedback will be accumulated as training material for system evolution. This means MGI is building a closed-loop ecosystem where "devices are the entry point, data is the fuel, and Agents are the operating system."
In the global AI for Bio race, model capabilities are rapidly being "leveled." OpenAI, Google, Anthropic, and domestic players like DeepSeek and Tongyi are all providing powerful foundational models at extremely low cost. However, the decisive factor for Physical AI lies precisely in the "physical half" that models cannot reach: real equipment, real wet-lab experiments, real failure cases, and real expert experience. These are capabilities that cannot be downloaded, are difficult to distill, and can only be developed through repeated laboratory work, constituting MGI's deepest moat.
According to Grand View Research data, the global laboratory automation market reached a size of $8.27 billion in 2024 and is projected to reach $18.39 billion by 2033, with a compound annual growth rate of 9.3%. Within this rapidly expanding market, agent systems "capable of truly entering the physical world to execute experiments" will occupy the top of the value chain. For capital markets, in an era of model parity, companies that possess the "real-world side" hold the power to define the rules of the next-generation competition. Furthermore, the pace of converting technological achievements into commercial revenue and the depth of collaboration with leading pharmaceutical companies and research institutions remain key metrics requiring ongoing follow-up.
Comments