Uncertainty-Aware Non-Prehensile Manipulation with Mobile Manipulators under Object-Induced Occlusion
Jiwoo Hwang, Taegeun Yang, Jeil Jeong, Minsung Yoon, Sung-eui Yoon
AI summary
Problem
Non-prehensile manipulation using only onboard sensors suffers from object-induced occlusion, which creates dangerous blind spots and leads to collisions with unexpected obstacles. Prior methods lack mechanisms to actively reduce these perceptual ambiguities during manipulation.
Approach
The authors propose CURA-PPO, a reinforcement learning framework that uses a Distributional Collision Estimator to predict collision risk and uncertainty as a probability distribution. By incorporating these metrics as intrinsic penalties, the policy learns to maneuver strategically to reduce occlusions while completing its task.
Key results
- Formulated non-prehensile manipulation under occlusion as an uncertainty-aware decision-making problem
- Developed CURA-PPO with a Distributional Collision Estimator to extract risk and uncertainty for active perception
- Achieved up to 3× higher success rates than baselines across varying object sizes and obstacle configurations
- Demonstrated learned active perception behaviors that proactively resolve sensor occlusions during manipulation
Why it matters
Provides a practical, perception-aware solution for safe autonomous manipulation in cluttered environments using only local onboard sensing.
Abstract
Non-prehensile manipulation using onboard sens- ing presents a fundamental challenge: the manipulated object occludes the sensor’s field of view, creating occluded regions that can lead to collisions. We propose CURA-PPO, a reinforcement learning framework that addresses this challenge by explicitly modeling uncertainty under partial observability. By predicting collision possibility as a distribution, we extract both risk and uncertainty to guide the robot’s actions. The uncertainty term encourages active perception, enabling simultaneous ma- nipulation and information gathering to resolve occlusions. When combined with confidence maps that capture observation reliability, our approach enables safe navigation despite severe sensor occlusion. Extensive experiments across varying object sizes and obstacle configurations demonstrate that CURA-PPO achieves up to 3× higher success rates than the baselines, with learned behaviors that handle occlusions. Our method provides a practical solution for autonomous manipulation in cluttered environments using only onboard sensing.