Whole-Body Model-Predictive Control of Legged Robots with MuJoCo
John Zhang, Taylor Howell, zeji yi, Chaoyi Pan, Guanya Shi, Guannan Qu, Tom Erez, Yuval Tassa, Zachary Manchester
AI summary
Problem
Model-based control for legged robots typically relies on custom, hard-to-reproduce dynamics models and solvers, creating a high barrier to entry and slowing community adoption compared to simulation-based reinforcement learning.
Approach
The authors implement a standard iterative LQR (iLQR) algorithm that leverages the MuJoCo physics engine for forward dynamics and finite-difference approximations for derivatives, paired with an interactive GUI for real-time parameter tuning on hardware.
Key results
- Real-time whole-body MPC deployment on quadruped and humanoid hardware
- Successful sim-to-real transfer for dynamic locomotion and bipedal walking
- Interactive GUI for real-time parameter tuning and behavior visualization
- ~30% tracking improvement using time-varying LQR feedback
Why it matters
Lowers the barrier to entry for real-world model-based control research, enabling rapid prototyping and broader adoption of whole-body predictive control on legged hardware.
Abstract
We demonstrate the surprising real-world effec- tiveness of a very simple approach to whole-body model- predictive control (MPC) of quadruped and humanoid robots: the iterative linear-quadratic regulator (iLQR) algorithm with MuJoCo dynamics and finite-difference approximated deriva- tives. Building upon the previous success of model-based behav- ior synthesis and control of locomotion and manipulation tasks with MuJoCo in simulation, we show that these policies can easily generalize to the real world with few sim-to-real consid- erations. Our baseline method achieves real-time MPC while leveraging whole-body dynamics and collision detection on a variety of hardware experiments, including dynamic quadruped locomotion, a quadruped walking on two legs, and full-sized humanoid bipedal locomotion. Additionally, our GUI system enables users to interactively update robot behavior in real- time on the robot hardware, making task-specific parameter tuning easy and intuitive. Our code and experiment videos are available online at: https://johnzhang3.github.io/mujoco ilqr.