Accessibility Field lets robots see tool-reachable surfaces in real-time LiDAR

A team of robotics researchers led by Bradley Scott at the University of Canterbury has introduced the Accessibility Field, a real-time system that determines which surfaces in a LiDAR scan are physically reachable by a specific tool. Published on arXiv (2608.06412), the work tackles a fundamental gap: standard 3D reconstruction and visibility estimation don't tell a robot whether a gripper, sensor, or spray nozzle can actually touch a surface. The method runs entirely on GPU, checking each surface point against precomputed geometry kernels for a tool at multiple approach angles, while a scan-centric Truncated Signed Distance Field (TSDF) updates only relevant voxels, making it efficient even for sparse, non-repetitive sensors like the Livox Mid-360.

Unlike prior geometric accessibility methods that require complete meshes and fixed bases, this approach is tool-agnostic, needs no pre-existing scene model, and works on edge hardware like the Jetson Orin. In experiments on synthetic objects and mature Pinus radiata tree models, the Accessibility Field achieved F1=90.8 versus 69.8 for a Hidden Point Removal visibility baseline. It also correctly identified 56.8% of pine branch surfaces as inaccessible to a pruning tool despite being fully visible from the sensor, highlighting that visibility is a poor proxy for accessibility. This is touted as the first method to estimate per-point surface accessibility in real time from streaming sparse LiDAR without a fixed base frame, opening the door for mobile manipulators to plan tool movements on the fly.