How to use Codex to train a self-righting robot
Codex built a MuJoCo and PPO training flow for a two-servo robot; its maker printed the body and demonstrated real self-righting with a known tilt limit.
homemadegarbage.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Build a MuJoCo and Gymnasium PPO training flow for my two-servo self-righting robot from these CAD files, masses, inertias, and joint axes. Respect the PWM servo speed limits using incremental angle actions. Evaluate fallen poses in simulation, export the policy for ATOM Matrix, and wait for my assembly and physical transfer checks.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- お父ちゃん (@H0meMadeGarbage)
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The trained policy transferred to the two-servo ATOM Matrix robot and produced repeated physical self-righting, with documented Euler-angle behavior at steep tilts.
- Human involvement
- The maker designed the body, supplied joint and servo constraints, assembled the electronics, and judged simulation-to-hardware behavior.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- The maker supplied CAD and physical parameters, then clarified the robot’s joint axes for Codex.
- Codex implemented a MuJoCo and PPO learning flow with incremental servo-angle actions.
- The maker printed and assembled the robot and demonstrated the learned policy standing up on real hardware.
HomeMadeGarbage designed a small two-servo robot and asked Codex to build the reinforcement-learning setup for getting it upright after a fall. The maker supplied Fusion geometry, mass and inertia information, and explicitly identified the rotation axes when the unusual mechanism was not inferred correctly.
Codex prepared a MuJoCo model, a Gymnasium environment, and a PPO training flow. The owner specified ordinary PWM servos and requested slow, dependable motion. The resulting actions were changes in target angle rather than abrupt target positions, reflecting those actuator limits.
After the simulation learned a usable policy, the maker printed the body and assembled an ATOM Matrix controller, two servos, and a battery. The trained policy ran on the physical robot, which repeatedly stood up after being knocked over. The public repository provides the simulation, printed parts, and embedded implementation.
There was a visible limit: Euler-angle readings could behave badly beyond a steep tilt, making some attempts less smooth. The author left that behavior in this demonstration and pointed to a quaternion-based correction in later work. Codex built the learning machinery; PPO trained the policy, while the human defined mechanics, assembled hardware, and assessed the transfer.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Maker blog · お父ちゃん (@H0meMadeGarbage)
Frequently asked questions
Did Codex itself learn the movement?
Codex built the environment and training flow; PPO trained the policy.
What required manual clarification?
The maker had to specify the unusual mechanism’s rotation axes.
What limitation remained in the demo?
Euler-angle behavior at steep tilts made some attempts less smooth.
More details
About this account
The maker supplied CAD and physical parameters, then clarified the robot’s joint axes for Codex. Codex implemented a MuJoCo and PPO learning flow with incremental servo-angle actions.
Firsthand Japanese maker account and linked SelfRisingRobot repository opened. Policy behavior was maker-demonstrated, not independently reproduced here. Euler-angle limitation preserved; later correction is not attributed to this run. Suggested prompt written by Agents In Use; no original task prompt was published in the verified account.
Credit: お父ちゃん (@H0meMadeGarbage) — Maker blog
Source published
Added to Agents In Use
Source checked
Tools and topics
Codex · MuJoCo · Stable-Baselines3 · Fusion · ATOM Matrix
reinforcement learning · sim-to-real · servo control
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.