Project 01

Controlling the Legs of a 12t Walking Excavator Using Reinforcement Learning

A reinforcement learning controller for a 12-ton Menzi Muck walking excavator. Mixed terrain of slopes and hills constructed in RaiSim simulation, with robustness increased via domain randomization and random initialization. PPO algorithm used for training, validated on real hardware.

Semester Project at ETH Zurich, Robotics Systems Lab (RSL)
Supervisors: Dr. Pascal Egli, Dr. Julian Nubert, Prof. Marco Hutter

Training

PPO training pipeline in RaiSim — controller network with PID, trained on parametrized terrain

Trained a PPO-based RL controller in the RaiSim physics simulator for a 12-ton Menzi Muck walking excavator
Constructed parametrized terrain with varying slopes, hills, and surface roughness
Increased robustness via domain randomization and random state initialization

Simulation

Flat terrain locomotion

Rolling hills

Steep slope traversal

Rough rocky terrain

Real-World Validation

Sim-to-real transfer — the real 12-ton Menzi Muck walks autonomously on gravel using the RL-trained controller

← Back to all projects