Multiple-Hypothesis Path Planning with Uncertain Object Detections

TL;DR

This paper introduces a multiple-hypothesis path planner that uses uncertainty reasoning and real-time graph updates to improve navigation success in obstacle-dense environments, outperforming traditional methods.

Contribution

It presents a novel path planning approach that incorporates evolving uncertainty and multiple hypotheses for safer navigation in complex environments.

Findings

Navigation success rate increased from 20% to 75% in challenging environments.

Evolving, range-based uncertainty improves obstacle inference accuracy.

Multiple hypotheses enable better safety and efficiency trade-offs.

Abstract

Path planning in obstacle-dense environments is a key challenge in robotics, and depends on inferring scene attributes and associated uncertainties. We present a multiple-hypothesis path planner designed to navigate complex environments using obstacle detections. Path hypotheses are generated by reasoning about uncertainty and range, as initial detections are typically at far ranges with high uncertainty, before subsequent detections reduce this uncertainty. Given estimated obstacles, we build a graph of pairwise connections between objects based on the probability that the robot can safely pass between the pair. The graph is updated in real time and pruned of unsafe paths, providing probabilistic safety guarantees. The planner generates path hypotheses over this graph, then trades between safety and path length to intelligently optimize the best route. We evaluate our planner on randomly generated simulated forests, and find that in the most challenging environments, it increases the navigation success rate over an A* baseline from 20% to 75%. Results indicate that the use of evolving, range-based uncertainty and multiple hypotheses are critical for navigating dense environments.