13 upvotes
New pipeline lets LLMs control robots via generated code
What is CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents?
CaPtain turns a pretrained language model into a robot controller by first fine-tuning it to emit Python code that talks to the robot, then polishing the code with reinforcement learning. The result is a policy that runs as ordinary code instead of a black-box network.
Who it is for
Robotics engineers and LLM hobbyists can reuse existing language models to program manipulators without writing low-level control loops.