KEPT: Conformal No-Go Zone Planning for Degradation-Aware Navigation in Confined Spaces

Eric Minwoo Kim and Jong-Kook Kim submitted a preprint proposing KEPT: a navigation method for confined spaces such as tunnels and mines. It uses conformal prediction to calibrate no-go zones for map walls and adjusts margins according to a label-free degradation-level estimate, keeping coverage near the target as degradation varies; the planning and control layers impose braking-distance conditions on the no-go zones at planning time and on each command, using measured tracking residuals as the sole margin without modifying the no-go zones. Validation has two parts: the no-go zones are tested on smoke-recorded data from a ground robot, and closed-loop tests are simulated quadrotor flights in a fog-filled tunnel, compared with a standard deployment. The authors report that in closed-loop simulation the method significantly reduces entries into calibrated no-go zones, and the number of goals reached is higher than standard deployment in both calibration cases, with a combined gain of 3.2 percentage points, though the confidence intervals for each case include zero; on held-out recorded data, per-class no-go zone coverage stays near the target.

AI 根据报道生成 · 14 days ago更新