Hrobot is a 230 mm, 20-DoF humanoid robot built from Dynamixel XL330 servos. Its walking, get-up, and rough-terrain policies are trained with reinforcement learning (PPO in mjlab, MuJoCo physics) and run directly on an ESP32-S3 at 50 Hz, using only the IMU and the servo encoders. Each policy is a small MLP that fits in 63 kB. The videos show the policies in MuJoCo physics. The open firmware is verified against ONNX Runtime on a PC but has not yet been tested on the physical robot.

roboticshumanoidreinforcement-learningsim-to-realembedded

Looking forFeedback

About

Dejan Batanjac is an AI Researcher and PhD Student based in Niš, Serbia, working as an independent researcher. Dejan trains humanoid robots to walk using reinforcement learning in MuJoCo, and uses PyTorch, Claude, and Hugging Face. Dejan publishes datasets, models, and Spaces on Hugging Face under the handle dejanb, including a reinforcement learning model for robot firmware.