BACON: A Body-Aware Framework for Parameterized Collision Detection Acceleration

Le résumé fourni par la source

The escalating intricacy of robotic systems has created an urgent need for customizable hardware platforms tailored to diverse robotic applications. Concurrently, the rapid iteration of robotic deployment scenarios demands an agile and scalable hardware development methodology. A critical but often overlooked module in agile hardware design is collision detection, a fundamental module in robot accelerators that is essential for safety and control. In this work, we present BACON, a hardware collision detection framework that generates high-precision, high-efficiency accelerators for rigid-body robots with arbitrary degrees of freedom and structures. BACON streamlines the design process with a parameterized template that can be rapidly tuned for new scenarios. Our framework introduces three key innovations: a dynamic-static decoupled design for efficient detection in highly dynamic environments; a Monte Carlo tree search-based optimization strategy that automatically tailors parameters for diverse robotic morphologies and scenarios; and a body-aware scheduler that balances workloads and eliminates redundant computations via early exit mechanisms. When validated on six robotic systems and implemented on Xilinx VCU118 and ZU15EG FPGA, BACON demonstrates an average speedup of 17.85× over CPUs and 2.62× over GPUs in collision detection tasks, highlighting its capability to enable scalable, high-performance robotic accelerators for a new generation of agile robotic systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.