Researchers from and present a GPU implementation of Hamilton-Jacobi reachability, a method that computes the set of states from which a robot cannot avoid an unsafe outcome. Historically this was computed offline; the authors say their solver is fast enough to run inside the control loop, which could serve as a formal safety check on learned policies.
On a workstation GPU it computes a three-dimensional benchmark case at 112 Hz and a four-dimensional pursuit-evasion case at 61 Hz, 5.9 to 17.5 times faster than a JAX reference at matched precision. On an NVIDIA Jetson AGX Thor embedded board it reaches 66 Hz on the three-dimensional case. A fleet extension ran at 4.0 ms per step in a simulated wildfire response with six quadcopters.
The results are benchmark timings and a simulation. The paper does not report the solver running a physical robot's safety layer.
Limits
Benchmark and simulation results only; no physical closed-loop deployment reported; not peer reviewed.