BioMARS: A Multi-Agent Robotic System for Autonomous Biological Experiments

• Biological protocols demand adaptive decisions and multi-stage coordination, but existing automation often requires manual oversight and lacks flexibility for unexpected deviations.Source: §1 · p. 2
• Although LLMs have been applied to chemical experimentation and biological protocol generation, their integration with physical robotic systems for biological execution remains underexplored.Source: §1 · p. 2
• BioMARS is a dual-arm robotic platform orchestrated by LLMs and VLMs for autonomous biological experiment design, planning, and execution.Source: §1 · p. 2
• BioMARS demonstrates five capabilities: literature-based protocol design, robotic translation and execution, keyframe error detection, cell culturing, and optimization using historical data.Source: §1 · pp. 2–3

2.1 Architecture of BioMARS system

• The Biologist, Technician, and Inspector agents respectively synthesize constraint-aware procedures, translate plans into robotic controls, and detect anomalies that can prompt replanning or user notification.The modular platform supports interactive experiment monitoring and modification through a web interface.Source: Figure 1 · p. 3, §2.1 · p. 4

2.2 Protocol Synthesis under Environmental Constraints

• DeepSeek-R1+WC+KC achieved consistent protocol scores of 5, including a HepG2 medium-replacement procedure with specified reagents, conditions, and handling steps.The evaluation covered 70 queries across 10 procedural categories and seven cell lines.Source: Figure 2 · p. 5, §2.2 · p. 6

2.3 Protocol-to-Code Translation for Robotics

• Across 300 protocol steps, the GPT-4o CodeGenerator-plus-CodeChecker pipeline achieved 96.4% instruction-matching accuracy, versus 92.4% for the single-module baseline.The CodeChecker also corrected parameter mismatches, enforced range constraints, and removed superfluous instructions.Source: Figure 3 · p. 7, §2.3 · pp. 7–8

2.4 Hierarchical VLM-Based Error Detection

• ViT keyframe detection achieved 88.7% F1 and 94.0% recall; VLM semantic validation reached 95.7% precision and reduced false positives from 8.7% to 1.4%.ViT mean inference latency was 0.3066 s, 91.9% faster than GPT-4o at 3.7960 s.Source: Figure 4 · p. 8, §2.4 · p. 9

2.5 Integrated Biological Experiment Design

• BioMARS matched or exceeded manual biological outcomes while reducing hands-on time by 90% and improving reproducibility without detectable morphological abnormalities.The comparison covered automated and manual cell passaging; HeLa and Y79 samples showed 12–18% lower variability with BioMARS.Source: Figure 5 · p. 10, §2.5 · pp. 10–11

2.6 Biological Optimizing Capability

• With 10 prior data points, DeepSeek-R1 reached a final pigment score of 0.5913, exceeding GPT-4o (0.4344; +15.8%) and Bayesian optimization (0.3130; +28.5%).Without prior knowledge, DeepSeek-R1 performed comparably to its prior-informed case; output-score standard deviation was 0.2366, versus 0.2447 for GPT-4o and 0.2785 for Bayesian optimization.Source: §2.6 · pp. 11–13, Figure 6 · p. 12

3 Discussion

• BioMARS integrates LLM- and VLM-based reasoning with multimodal perception and robotic control for autonomous experiment design, planning, and execution.The authors say this integrated system generates reproducible, high-quality outcomes.Source: §3 · p. 13
• BioMARS is robust in standard cell-culture tasks, but atypical conditions, critical pipetting or centrifugation steps, and unexpected deviations remain challenging.Dependence on existing online procedures also limits adaptive parameter tuning across laboratory contexts.Source: §3 · p. 13
• Consistent procedural replication supports reproducibility and quality control; automation may reduce time, material use, labor, and operational expenses as production scales.The authors identify drug discovery and cell-model production as relevant applications.Source: §3 · p. 13

4.1 Construction of LLM-based and VLM-based agents

• The Biologist uses retrieval and checking to build constrained protocols; the Technician validates executable pseudo-code; the Inspector combines visual anomaly detection with VLM interpretation.The Inspector’s monitoring is hierarchical, pairing geometric detection with contextual interpretation.Source: §4.1 · p. 14

4.2 Web application

• The Flask web interface supports visualization, alerts, and interactive task control, while bidirectional WebSocket communication synchronizes agents, interface, and hardware.The Python backend orchestrates agents and robotic workflows.Source: §4.2 · p. 14

4.3 Modular Robotic System for Automated Protocols

• A dual-arm platform combines modular handling stations, a centrifuge, and CO incubator; ROS actions and 11 base API functions abstract hardware controls and support plug-and-play integration.The workstations include lid actuators and tilting mechanisms for container manipulation.Source: §4.3 · pp. 14–15

4.4 Architecture of Hierarchical Error Detection

• Visual screening compares processed frame features with task-specific references using adaptively selected α-quantile thresholds; detected anomalies trigger semantic validation and reports to users and the Biologist Agent.Semantic validation pairs the original frame’s language interpretation with task-specific constraints.Source: §4.4 · p. 15

Author contributions:.

• Y.Qiu led system development and technical validation; contributors covered robotics, workflows, mechanical design, biological experiments, statistical analysis, protocols, and project supervision.The listed authors collectively wrote the manuscript.Source: Author contributions · p. 15

Competing Interests.

• The authors declare no competing interests.Source: Competing Interests · p. 15

Data availability.

• Supplementary Information provides examples of experiments discussed in the text.Source: Data availability · p. 15

Code availability.

• The implementation is available at the stated GitHub repository.Source: Code availability · p. 15

A.1 Core Algorithm Implementation

• The Code Generator maps the checked workflow to the 11-action API, and the Code Checker validates logic and interfaces before execution.The algorithm returns experiment results and an execution log with errors.Source: §A.1 · p. 19
• Literature search and RAG acquire protocol knowledge, which the Knowledge Checker reorganizes before workflow generation.Source: §A.1 · p. 19
• The Workflow Generator combines checked knowledge, the query, and experimental environment; the Workflow Checker adapts the draft to that environment.Source: §A.1 · p. 19
• During each validated action, BioMARS monitors execution with ViT keyframe detection and invokes VLM semantic analysis when the detector raises a warning.A detected semantic error is logged, alerts the user, aborts the action, and sends the workflow for replanning.Source: §A.1 · p. 19

B.1 10-Query Biological Task Set

• The 10-query set applies the same questions to HeLa, HUVEC, HepG2, DC2.4, Y79, K562, and CHO.Only the cell-type name varies between versions of each question.Source: §B.1 · p. 20
• The questions cover routine cell handling, including resuscitation, passaging, medium changes, freezing, and storage.Source: §B.1 · p. 20
• The set also asks about seeding density, metabolic activity, apoptosis, cryopreservation solution, proliferation, and 3D cell-sphere culture.Source: §B.1 · p. 20

Score 5

• The DeepSeek-R1+WC+KC response to a HepG2 medium-change question is characterized as very detailed and biologically accurate.Source: Score 5 · p. 20
• The procedure aspirates 10 ml of old medium, rinses with 10 ml pre-warmed PBS, then adds 10 ml pre-warmed culture medium.The PBS and medium are pre-warmed to 37°C.Source: Score 5 · p. 20
• The example specifies 37°C and 5% CO conditions and horizontal agitation, alongside precise reagent volumes.Source: Score 5 · p. 20

Score 4

• The procedure specifies 10 ml PBS and fresh-medium volumes and gives operationally correct steps, but omits critical medium composition details such as DMEM + 10% FBS.It repeats a PBS wash and swirls the fresh medium in a figure-eight for 15 seconds before returning ContainerA to the incubator.Source: Score 4 · p. 21

Score 3

• Score 3 describes a HUVEC cryopreservation response with logical errors, including inappropriate incubator storage and parameter problems such as centrifugal force.The example divides cells between two containers and places them in the incubator for cryopreservation.Source: Score 3 · p. 21

Score 2

• Score 2 describes a HepG2 medium-change response that adds irrelevant passaging steps, including trypsin digestion and centrifugation.The assessment notes that medium change does not require passaging.Source: Score 2 · p. 22

Score 1

• Score 1 describes a HeLa resuscitation response whose steps are judged completely unclear.Source: Score 1 · p. 22

C.1 Functional Interface Specifications

• Table C1 is titled as a function list for cell culture operations.Source: Table C1 · p. 23
• The table's stated scope is cell culture operations.Source: Table C1 · p. 23
• The functional interface specifications include a table labeled C1.Source: Table C1 · p. 23

D.1 Error Scenario

• Missing or detached tips, absent dishes or tubes, and unopened or disconnected containers occur across liquid-removal and addition steps.The scenarios also include failure to lift the platform during aspiration.Source: §D.1 · pp. 24–25
• Liquid ejection can overflow when the platform is not lifted; other listed failures include rotor or tube problems and tip loss during resuspension or cell pipetting.The scenarios include a rotor absent before tube insertion or after centrifugation, and no tubes prepared.Source: §D.1 · p. 25

E.1 Hardware

• The workspace places custom devices and standard laboratory equipment within reach of both arms, alongside dedicated dish, reagent-bottle, and centrifuge-tube stations.The setup includes a three-position reagent-bottle station, Petri dish handling and storage, a capper/decapper, waste container, incubator, and centrifuge.Source: Figure E1 · p. 26

F.1 Web-Based Control Panel Design

• The web panel combines a navigation menu with a content area for robot monitoring and control, emergency stops, command input, activity feedback, and knowledge-base document management.The home dashboard also provides shortcuts, while its status overview and activity timeline are described as under development.Source: §F.1 · pp. 27–28