SafeHarness: Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation

Forum topic · 小凯 · 2026-09-19

Summary

This paper examines the safety of coding agents—a promising paradigm in which a language model writes robot controllers as programs—under obstacle-avoidance constraints in robot manipulation. The authors find that agents typically collide with obstacles despite reasoning about them and receiving explicit prohibitions in prompts, revealing that the failure lies in planning: safety constraints never become priorities. During path phases, models lack the concept of obstacle-avoiding routes and do not replan when chosen paths are infeasible; during contact moments, they fail to recognize that contact execution is also constrained. To address this, the authors propose SafeHarness, adding two obstacle-aware safeguard mechanisms: obstacle-aware path planning, which grounds obstacles as bounding boxes and plans/validates waypoint paths before execution, and obstacle-aware contact execution, which selects contact points that avoid obstacles. SafeHarness achieves 71.9% task success and 87.5% obstacle avoidance, surpassing prior state-of-the-art by 6.5% and 27.0% respectively, and outperforming the same agent without the safety framework by 2.3x and 1.5x. The work is available on arXiv (2609.20822) with authors from a team including Bingxin Xu, Yuzhang Shang, Zhen Dong, and Emilio Ferrara.

Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation

Paper Overview

Field: NLP
Authors: Bingxin Xu, Yuzhang Shang, Zhen Dong, Emilio Ferrara
Published: 2026-09-17
arXiv: 2609.20822

Summary

Coding agents have emerged as a promising paradigm for robot manipulation: a language model writes the robot controller as a program, and agents built in this way now operate robots without robot-specific training. Whether this paradigm is also safe, however, has not been asked. The authors evaluate coding agents under a safety constraint, where each task pairs a manipulation goal with an obstacle the robot must not touch. The agent pursues the goal but collides with the obstacle in most cases, treating task completion as its sole objective while neglecting safety. The agent reasons about the obstacle in its traces, and the prompt already forbids touching it, so neither perception nor instruction is at fault; the fault lies in the planning, where the stated constraint never becomes a priority.

By decomposing manipulation into path phases and contact-rich moments, the authors localize the failure. In path phases, the model cannot prioritize safety constraints—it lacks the concept of an "obstacle-avoiding path" and does not replan when the chosen path is infeasible. At contact moments, it fails to realize that contact execution is subject to the same constraint.

The Proposed Solution: SafeHarness

To close this gap, SafeHarness equips the model with two obstacle-aware safety constraint prioritization safeguards:

1. Obstacle-aware path planning: obstacles are grounded as bounding boxes, and candidate paths are drawn over them as waypoint sequences. The agent pre-plans the path, validates it, replans if necessary, and only then executes. 2. Obstacle-aware contact execution: contact locations are chosen so that the contact itself avoids obstacles.

Results

SafeHarness achieves 71.9% task success and 87.5% obstacle avoidance, exceeding the previous state-of-the-art by 6.5% and 27.0% respectively. These results are 2.3x and 1.5x those of the same agent without the safety framework.

---

*Auto-collected on 2026-09-19.*

SafeHarness: Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation

Summary

Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation

Paper Overview

Summary

The Proposed Solution: SafeHarness

Results

Tags