Uncertainty-Aware Non-Prehensile Manipulation with Mobile Manipulators under Object-Induced Occlusion

Jiwoo Hwang, T. Yang, Jeil Jeong, Minsung Yoon, Sung‐Eui Yoon

TL;DR

Researchers have developed CURA-PPO, a reinforcement learning framework that allows mobile robots to push large objects safely, even when those objects completely block the robot's "eyes" (onboard sensors). By treating collision risk as a probability distribution rather than a single number, the robot learns "active perception" behaviors—like peeking around a box—to spot hidden obstacles, resulting in up to a 3x higher success rate than standard methods.

What Problem Does This Paper Address?

In robotics, non-prehensile manipulation (pushing or sliding an object rather than grasping it) is essential for moving items that are too heavy, bulky, or awkwardly shaped for a robot's grippers. However, there is a major catch: if a mobile robot uses a base-mounted sensor (like LiDAR) to see, the very object it is pushing often creates a massive blind spot directly in its path.

Imagine trying to push a large refrigerator through a cluttered room while only being able to see through a camera mounted at your shins. You can't see what's directly behind the fridge. Existing algorithms often struggle with this "object-induced occlusion," leading to collisions with "unforeseen" obstacles like fallen boxes or people entering the workspace. The challenge is to navigate safely using only local, partial information without the luxury of a "god's eye view" from overhead cameras.

Methods

The authors introduce CURA-PPO (Collision Uncertainty-Risk Aware PPO), which combines several clever components:

  1. Distributional Collision Estimator (DCE): Instead of simply predicting if a collision will happen, the DCE predicts a distribution of possible collision outcomes. This allows the robot to distinguish between "I'm pretty sure it's safe" (low variance) and "I have no idea if it's safe" (high variance).
  2. Risk vs. Uncertainty: The framework extracts two signals from the DCE:
  3. Risk (Mean): Encourages the robot to avoid known obstacles.
  4. Uncertainty (Variance): Acts as an "intrinsic cost" that motivates the robot to move in ways that gather more information (active perception).
  5. Confidence Maps: The robot maintains a local map where sensor data "decays" over time. If the robot hasn't seen a patch of floor recently because a box is blocking it, the confidence in that area drops, signaling the robot to go check it out.
  6. Reinforcement Learning (PPO): These risk and uncertainty signals are baked into the reward function of a Proximal Policy Optimization (PPO) algorithm, training the robot to balance reaching its goal with the need to stay "informed" and safe.

Datasets

The researchers evaluated their method using NVIDIA Isaac Sim, a high-fidelity robotics simulator.

  • Hardware Setup: A simulated ClearPath Ridgeback omnidirectional base equipped with a 7-DOF Franka Emika research arm and a 2D LiDAR.
  • Scenarios: They tested across two main environments:
  • Uniform: Obstacles are scattered randomly.
  • Adversarial: Obstacles are intentionally placed along the direct path to the goal, specifically hidden behind the manipulated object.
  • Scale: The evaluation used 10,000 test cases across varying object sizes (0.5m to 1.0m cuboids).

Key Insights and Ideas

  • Uncertainty as a Feature, Not a Bug: The core "aha moment" is that uncertainty shouldn't just be managed—it should be used as a compass. By penalizing high uncertainty, the robot naturally learns to perform lateral "zig-zag" maneuvers to peek around the object it's pushing.
  • Kinematic Redundancy is Key: A mobile manipulator (base + arm) has more "degrees of freedom" than it needs for simple pushing. The robot learns to use this extra flexibility to shift its base for a better view while keeping the arm steady on the object.
  • The "Fading" Memory: Using confidence maps with temporal decay is crucial. It prevents the robot from assuming a path is clear just because it saw it was empty 10 seconds ago—a vital feature for dynamic environments where things move.

Key Results

  • Massive Success Gains: In the most difficult "Adversarial" scenarios with large (1.0m) objects, the baseline policy's success rate plummeted to 22.9%. CURA-PPO maintained a success rate of 72.4%—a 3x improvement.
  • Proactive Safety: Qualitative analysis showed that while baseline robots blindly pushed forward into collisions, CURA-PPO robots executed purposeful lateral movements to reveal hidden obstacles before they became a threat.
  • Robustness to Size: As the object size increased (and thus the occlusion grew), the performance gap between CURA-PPO and standard methods widened significantly, proving the method's scalability.

Limitations and Future Work

  • Sensor Modality: The current work relies on 2D LiDAR. While effective, the authors note that 3D sensors or RGB-D cameras would be necessary for more complex, real-world environments with low-profile or overhanging obstacles.
  • Efficiency Trade-off: Active perception maneuvers (zig-zagging) are inherently less "efficient" in terms of path length than a straight line. Future work could focus on optimizing the balance between the "cost of looking" and the "cost of moving."
  • Sim-to-Real: The study was conducted entirely in simulation. The next major step is deploying this on physical hardware to see how sensor noise and real-world physics affect the learned active perception behaviors.

Summary updated

Abstract

Non-prehensile manipulation using onboard sensing 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 manipulation 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 3X 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.
BibTeX
@misc{hwang2026uncertainty,
  author = {Jiwoo Hwang and T. Yang and Jeil Jeong and Minsung Yoon and Sung‐Eui Yoon},
  title = {{Uncertainty-Aware Non-Prehensile Manipulation with Mobile Manipulators under Object-Induced Occlusion}},
  year = {2026},
  eprint = {2602.01731},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.01731},
}
Figure