← Back ICRA 2026

Semantically-Driven Deep Reinforcement Learning for Inspection Path Planning

Grzegorz Malczyk, Mihir Kulkarni, Kostas Alexis

PDF

AI summary

Key figure (auto-extracted from paper)

A semantics-aware deep reinforcement learning policy enables autonomous flying robots to efficiently inspect target objects while safely navigating unknown environments, successfully bridging the simulation-to-reality gap.

Semantic inspection Deep reinforcement learning Autonomous path planning Sim2real transfer Aerial robotics Ego-centric perception

Problem

Traditional exploration algorithms inefficiently cover entire environments without prioritizing relevant targets, while existing reinforcement learning methods lack semantic awareness and struggle to transfer to real-world robotic platforms.

Approach

The method trains an end-to-end deep RL policy using only local ego-centric occupancy maps, spatial visit history, and semantic-masked depth images to simultaneously learn collision-free navigation and targeted visual inspection.

Key results

Achieves up to 92.1% semantic surface coverage in simulated environments
Successfully transfers trained policy to real-world aerial robot flights
Boosts inspection coverage by 22.3% using spatial visit score maps
Generalizes across unknown scenes and unseen semantic classes without prior maps

Why it matters

Enables efficient, targeted autonomous inspection for industrial monitoring and disaster response by focusing computational and navigational resources on objects of substantive interest.

Abstract

This paper introduces a novel semantics-aware in- spection planning policy derived through deep reinforcement learning. Reflecting the fact that within autonomous informative path planning missions in unknown environments, it is often only a sparse set of objects of interest that need to be inspected, the method contributes an end-to-end policy that simultaneously per- forms semantic object visual inspection combined with collision- free navigation. Assuming access only to the instantaneous depth map, the associated segmentation image, the ego-centric local occupancy, and the history of past positions in the robot’s neighborhood, the method demonstrates robust generalizability and successful crossing of the sim2real gap. Beyond simulations and extensive comparison studies, the approach is verified in experimental evaluations onboard a flying robot deployed in novel environments with previously unseen semantics and overall geometric configurations.

Index terms

Aerial Systems: Perception and Autonomy Aerial Systems: Applications Reinforcement Learning