Communications of the ACM
Using LLMs to Code New Tasks for Robots
The researchers worked to automatically generate rich simulation environments and expert demonstrations by exploiting a large language models' grounding and coding ability.
Credit: MIT Computer Science and Artificial Intelligence Laboratory
A team that included Massachusetts Institute of Technology (MIT) researchers developed a tool that uses large language models (LLMs) to code new tasks for robots, which it then simulates.
GenSim features goal-directed and exploratory settings. GenSim breaks down each step necessary to complete the objective when in goal-directed mode and formulates new tasks when in exploratory mode.
The process, regardless of mode, requires an LLM to produce the task descriptions and the code needed for the simulation and refine the code using a task library.
The researchers determined that GenSim, which was pretrained on 10 tasks, produced 100 new behaviors on its own and was able to train robotic arms to execute tasks at a higher success rate than comparable methods.
MIT's Lirui Wang said they showed "that GenSim works in both simulation and the real world."
From MIT Computer Science and Artificial Intelligence Laboratory
View Full Article
Abstracts Copyright © 2023 SmithBucklin, Washington, D.C., USA
No entries found