CUDA-accelerated MPC slashes training time for drone racing AI
New algorithm cuts end-to-end latency for high-performance drone control.
Actor-critic model predictive control (AC-MPC) combines the long-horizon planning of MPC with the adaptive learning of reinforcement learning, but its differentiable MPC layer repeatedly solves optimization problems in both forward and backward passes, creating a severe latency bottleneck. Researchers from the University of Naples (Buo et al.) tackle this by introducing CA-AC-MPC, a CUDA-accelerated variant that parallelizes the optimization on GPU hardware, drastically cutting end-to-end execution time while preserving control performance.
Tested on an agile drone racing task, CA-AC-MPC achieves state-of-the-art lap times and near-limit dynamic behavior with markedly reduced training and inference time. The paper has been accepted for presentation at the 2026 International Conference on Unmanned Aircraft Systems (ICUAS 2026). This work demonstrates that GPU acceleration can make computationally intensive control algorithms feasible for real-time autonomous systems, opening the door to faster, more responsive drones and robots.
- Combines model predictive control with reinforcement learning in a single differentiable architecture
- CUDA parallelization reduces training and inference latency by offloading optimization to GPUs
- Achieves state-of-the-art lap times in agile drone racing benchmark
Why It Matters
Makes advanced real-time control practical for autonomous drones and robots, enabling faster, more agile maneuvers.