LeCAR-Lab/dial-mpc

Health v2 · maintenance only

Adoption not part of the score

1011 stars · 104 forks observed · 2026-10-20

What it is AI-extracted, prompt v1, taxonomy v1, 2026-09-07, confidence 0.95

DIAL-MPC is a sampling-based model predictive control framework for full-order, torque-level locomotion control of legged robots, using diffusion-style annealing. It is training-free, rolls out directly in Brax physics simulation, and includes sim-to-real deployment support for robots like the Unitree Go2.

Use cases

• control humanoid or quadruped robots at torque level without training
• run sampling-based MPC in JAX/Brax physics simulation
• deploy locomotion controllers to real Unitree Go2 robots via sim2real
• prototype whole-body robot control tasks like object reorientation
• compare diffusion-style annealing MPC against reinforcement-learning controllers

When to choose

• you need precise and agile legged robot control without training RL policies
• you want a plug-and-play MPC controller with minimal reward and dynamics model requirements
• you work with Unitree or similar legged platforms and need sim2real

When to avoid

• you need very low-latency control on non-GPU hardware
• your robot is not supported by Brax/MuJoCo simulation
• you prefer learning-based policies or classical reduced-order MPC

Facets

library · maturity active

simulation robotics machine-learning robotics simulation machine-learning python mpc model-predictive-control diffusion-annealing legged-robots torque-level-control training-free brax sim2real humanoid quadruped linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="LeCAR-Lab/dial-mpc")