Language models for laboratory robotics
My master’s thesis evaluates open-weight language models as high-level planners for laboratory robotics, in collaboration with NML Harada Lab at the University of Tokyo.
- Developed Python benchmark and evaluation scripts, plus C#/Unity components for model interaction, scene management, and robot API integration.
- Extended an existing digital twin to preview plans in Unity and execute approved plans on a physical robot.
- Compared 44 configurations to examine planning correctness, appropriate abstention, output stability, and failure modes.
- Scenarios
- 53
- Final runs
- 212
- Scenario executions
- 11,236
The benchmark measures command generation; the physical demonstration illustrates integration feasibility.