Reproduce “Occluding the Solution Space: Planner-Agnostic Adversarial Attacks on Tolerance-Aware Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Keke Tang, Tianyu Hao, Weilong Peng, Hao Jiang, Feng Wu, Peican Zhu, Jianmin Ji, Zhihong Tian
A framework for planner-agnostic adversarial attacks on robotic manipulation motion planning is presented, using strategically placed obstacles to occlude an entire task-level goal region. The method consistently achieves complete planning failure with fewer obstacles than baselines and demonstrates effectiveness in both high-fidelity simulation and real-world robotic deployments.
Key insights
- Adversarial attacks on manipulation tasks are more effective and realistic when targeting the occlusion of an entire feasible goal region, rather than merely obstructing a single path to an exact pose.
- Effective planner-agnostic attacks can be generated by exploiting intrinsic geometric and kinematic vulnerabilities (spatial bottlenecks) of the robot, without requiring internal knowledge or iterative queries to the victim planner.
- Kinematic bottlenecks represent a fundamental, shared physical vulnerability for diverse robotic planning approaches, including both classical sampling-based planners and modern end-to-end learned policies like VLAs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.