MuJoCo MPC (MJPC)

Google DeepMind's open-source tool for interactive, real-time model predictive control inside MuJoCo.

MuJoCo MPC (MJPC) is an interactive tool Google DeepMind open-sourced in 2022 that uses the MuJoCo simulator to run model predictive control — at every instant, simulating a short window into the future, optimizing an action sequence, executing only the first step, and repeating. It has built-in planners including iLQG, gradient descent, and Predictive Sampling; a user adjusts cost-function weights in a graphical interface and sees the robot's behavior change live. It doesn't train a neural network at all, which makes it well suited to quickly generating motion for quadrupeds, humanoids, or dexterous hands, and it's also commonly used as a model-based control baseline to compare against reinforcement learning.

ExamplePick a quadruped task in the MJPC interface, drag a target point, and the robot walks over to it via real-time online planning.

Also called

MJPC

Related

Model Predictive Control (MPC) · Sampling-based MPC · Iterative Linear Quadratic Regulator (iLQR) · MuJoCo (Multi-Joint dynamics with Contact) · DIAL-MPC (Diffusion-Inspired Annealing for Legged MPC)

Sources

mujoco_mpc GitHub
Predictive Sampling: Real-time Behaviour Synthesis with MuJoCo (arXiv)