Reproduce “cuNRTO: GPU-Accelerated Nonlinear Robust Trajectory Optimization” on OpenResearch

About this paper

Jiawei Wang, Arshiya Taj Abdul, Evangelos A. Theodorou

cuNRTO, a GPU-accelerated framework, addresses the high computational cost of Nonlinear Robust Trajectory Optimization (NRTO) by re-architecting its core components for parallel execution. The framework achieves speedups of up to 139.60x over CPU-based NRTO while maintaining robust constraint satisfaction across various robotic tasks.

Key insights

• Restructuring robust trajectory optimization problems to fully leverage GPU architectures, particularly by integrating SOCP projections directly into an on-device ADMM loop, yields substantial computational speedups.
• Minimizing CPU-GPU data transfers and maximizing GPU utilization through custom CUDA kernels for bottleneck operations like SOC projections are crucial for drastically improving the efficiency of robust optimization.
• Significant algorithmic and hardware acceleration can be achieved without compromising the quality of the robust solution or its ability to satisfy constraints, as evidenced by comparable objective values and maintained feasibility.

Summary by alphaXiv.