Empowering AI to Autonomously Design and Execute Biological Experiments: The Rise of Robotic Biologists and Laboratory Automation}
BioMARS system enables AI-driven autonomous design and execution of biological experiments, revolutionizing lab automation with improved accuracy, efficiency, and scalability.
Editor | ScienceAI
In biological research, repetitive experiments like cell culture and parameter optimization are time-consuming and prone to human error. Traditional automation (e.g., liquid handling robots) lacks flexibility—struggling with complex protocols and dynamic conditions, often requiring manual intervention to correct errors or optimize processes.
How can AI truly “understand” and autonomously perform biological experiments? This question has become a key bottleneck in accelerating scientific discovery.
Recently, a team from the University of Science and Technology of China’s Suzhou Institute of Advanced Technology introduced the BioMARS system (Biological Multi-Agent Robotic System). This system deeply integrates large language models (LLMs) and visual language models (VLMs) into robotic platforms, achieving end-to-end fully automated biological experiments!
The system not only matches or surpasses manual cell culture results in authoritative tests but also autonomously optimizes experimental parameters, significantly boosting efficiency and reproducibility.
Figure 1: System architecture and robot setup
What’s the breakthrough?
Three agents collaborate like human researchers—thinking and acting!
BioMARS innovatively adopts a layered multi-agent architecture, enabling AI “division of labor” to complete experiments:
Figure 2: Biologist Agent architecture and evaluation
1. Biologist Agent
Uses retrieval-augmented generation (RAG) to analyze literature and automatically synthesize executable protocols. For example, input “How to passage HeLa cells,” and it generates detailed steps compatible with lab constraints (e.g., dish inventory, pipette capacity), ensuring direct execution.
Figure 3: Technician Agent architecture and protocol conversion/execution performance
2. Technician Agent
Converts natural language protocols into robot pseudocode, supporting 11 basic operations (e.g., pipetting, centrifugation). Using a “CodeGenerator + CodeChecker” dual-module system, it automatically repairs logical errors (e.g., missing implicit steps), achieving 96.4% instruction accuracy, far exceeding single-module baselines (92.4%).
Figure 4: Inspector Agent overview and evaluation metrics
3. Inspector Agent
Uses visual transformer (ViT) and VLM for real-time monitoring of experiments, implementing layered error detection: initial quick assessment via ViT, followed by semantic validation with VLM (e.g., “culture dish not sealed”). In 23 common error scenarios, detection accuracy reaches 95.7%, with dual-stage detection reducing false positives by 83%, ensuring experiment robustness.
Results:
Cell culture quality comparable to manual, with significantly improved optimization capabilities!
Figure 5: Automated vs. manual cell passaging comparison
Cell passaging tasks in HeLa, Y79, and other cell lines show >92% cell viability post-operation, with CV reduced by 12-18%, demonstrating higher consistency. Operation time drops from 60 minutes to 5-8 minutes, a 90% efficiency increase!
Figure 6: iPSC-RPE optimization results
In iPSC-RPE differentiation, BioMARS adjusts 7 parameters (e.g., FGFRi concentration, centrifugation time). After 20 iterations, pigmentation score reaches 0.5913, surpassing Bayesian optimization (0.3130) and GPT-4o (0.4344), demonstrating strong reasoning and generalization!
Why is this revolutionary?
Language-driven + robot integration ushers in an “intelligent era” of lab automation!
BioMARS is more than a tool—it’s an “AI researcher”: it autonomously executes tasks via natural language interactions (e.g., “How to passage cells?”), supporting real-time human-machine collaboration. Its modular backend can integrate various lab hardware, reducing manual labor by 90%, and providing scalable solutions for drug discovery, regenerative medicine, and more. Future plans include enhancing anomaly handling and moving toward fully autonomous labs!
Paper link: https://arxiv.org/pdf/2507.01485
Online demo: https://github.com/AlexandreQ27/BioMARS