TurboMPC runs 58x faster GPU control for robots
58x faster than existing differentiable solvers—GPU-native MPC for robots.
TurboMPC, a new differentiable MPC solver from Gabriel Bravo-Palacios, Jianghan Zhang, and colleagues at MIT and Columbia University, is designed to run entirely on the GPU, addressing the growing need for parallel simulation and large-scale learning in robotics. The solver leverages sequential quadratic programming (SQP) with an alternating direction method of multipliers (ADMM) inner loop, implicit differentiation, and a co-designed JAX-CUDA implementation. It supports state and control inequality constraints, implicit integrators, cross-time-coupled costs, and slack variables, making it compatible with the expressive formulations required in modern robotics.
In simulations, TurboMPC achieved up to 15× speedup over state-of-the-art CPU solvers and an impressive 58× speedup over GPU-based differentiable solvers. The researchers validated it on constrained planning, humanoid imitation learning, and reinforcement learning with neural-network cost functions. Real-world deployment on a full-scale car for minimum-time racing showed that batched GPU-accelerated tuning of MPC parameters via Bayesian optimization produced significantly faster driving than a hand-tuned baseline. TurboMPC also demonstrated scalability by maintaining control over planning horizons exceeding 8,000 knot points. The code is open-sourced.
- GPU-native MPC solver combining SQP with ADMM and implicit differentiation, written in JAX-CUDA
- Up to 15× faster than CPU solvers and 58× faster than GPU differentiable solvers in benchmark tasks
- Deployed on a full-scale racing car; scaled to 8,000+ knot point planning horizons with real-time control
Why It Matters
GPU-accelerated MPC unlocks faster, more adaptive robot control for autonomous driving and legged locomotion.