Problem. Planners like EGO-Planner fly fast but offer no guarantee near moving pedestrians; safety filters that only brake sacrifice progress. The goal is the fastest safe go-around, not the safest stop.
Prediction. Constant-acceleration Kalman filter over pedestrian tracks; prediction margins calibrated with conformal prediction, checked by age-conditional coverage diagnostics.
Certification. Bernstein / de Casteljau exact continuous-time collision certificate over the B-spline trajectory, using outward-rounded interval arithmetic so the check is sound by construction.
Evaluation. 38-scenario MetaUrban benchmark, 100-seed A/B runs under matched perception: zero collisions across 304 certified flights vs. 9.2% for native EGO-Planner, with a proven ≤5% collision-probability bound at a 4.6% cost in goal-reach rate.
Tooling
Simulation and benchmarks: MetaUrban, ROS, EGO-Planner (cubic B-splines).
Analysis: Python, NumPy, conformal-prediction utilities; plots and diagnostics generated from the same experiment logs as the paper figures.
Earlier pipelines that still inform this work: Blender synthetic-data generation with domain randomization (Midea), RL training in RoboCasa / Isaac Gym / MuJoCo (KUKA).
Questions in progress
How tight can the certificate margins get before conservatism eats the go-around advantage?
Does age-conditional conformal calibration hold when pedestrian behaviour shifts between scenario families?
What is the cleanest way to expose a certified/uncertified decision to a human supervisor during teleoperation?
