Thermite for Robotics
Rewards and robot software checked by an independent test of success.
Reinforcement learners exploit any gap in their reward function: agents design the reward, and the judge trains each policy from scratch and evaluates task success with its own test. A goal campaign moves a robot's software from ROS 1 to ROS 2, checked by builds, launch tests and missions in simulation.
Robot reward design
Agents write and tune reward functions. The judge trains each policy from scratch in simulation and counts task successes with its own test, not with the agents' reward.
For Local AI · Startups
- Measures
- Task success rate
- Holds
- The judge's own test of success
- Hidden
- Seeds and dynamics the agents never see
ROS 1 to ROS 2 migration
Agents port the robot's packages to ROS 2. The judge builds them with colcon, runs the tests and launch tests, and drives missions in Gazebo; the run ends when every goal passes twice.
For Startups · Enterprises
- Measures
- Goals met, subsystem by subsystem
- Holds
- Tests, missions and topic rates hold
- Hidden
- Gazebo worlds the agents never see
Guides by audience
Twenty minutes to a first verdict.
Install it on one computer, or ask your own AI agent to:
Set up Thermite on this computer for me: follow https://usethermite.com/llm-setup.txt