DIAL-MPC

A training-free, sampling-based MPC for legged robots that borrows diffusion-style annealing to optimize whole-body motion in real time.

DIAL-MPC was proposed by Guanya Shi's group (LeCAR Lab) at Carnegie Mellon University in September 2024, and was a best-paper finalist at ICRA 2025. Real-time optimal control for legged robots usually requires a simplified model (such as a single rigid body) or a pre-specified contact schedule, because the full dynamics are high-dimensional and non-convex. Sampling-based MPC methods (such as MPPI) sample many candidate action sequences every control cycle, roll them out in a model, and average them by weight — but a single round of sampling is often noisy or gets stuck in a local solution. The authors identify a connection between MPPI and single-step diffusion denoising, and change it to iterate over multiple rounds the way a diffusion model does, annealing the noise level step by step: searching broadly at first, then converging to fine detail. It needs no training and no model simplification, doing real-time, torque-level control directly on the full-order dynamics.

ExampleOn a Unitree Go2 quadruped, DIAL-MPC performed precise, loaded jumps and trajectory tracking in real time; the paper reports 13.4x lower tracking error than standard MPPI and about 50% better performance than a reinforcement-learning policy on a climbing task, with no training at all.

Also called

Diffusion-Inspired Annealing for Legged MPC, Full-Order Sampling-Based MPC for Torque-Level Locomotion Control via Diffusion-Style Annealing

Related

Sampling-based MPC · Model Predictive Path Integral Control · Model Predictive Control · Legged Locomotion · Diffusion Model · Unitree Go2

Sources

Full-Order Sampling-Based MPC for Torque-Level Locomotion Control via Diffusion-Style Annealing (arXiv:2409.15610)
DIAL-MPC 项目主页（LeCAR Lab） (Chinese)

As of

2025