The authors create a robot manipulation benchmark where each task pairs a manipulation goal with an obstacle the robot must not touch, and they evaluate a coding agent—a language model that writes robot controller programs—on it.
Obstacle-aware harness improves safety of coding agents for robot manipulation
Authors show that language-model coding agents collide with obstacles by neglecting safety constraints during planning, and that a harness providing obstacle-aware route and contact plans reduces collisions by 27 percentage points.
Top University
Bingxin Xu · Yuzhang Shang · Zhen Dong · Emilio Ferrara
USC · UCF · UCSB
Research Digest··2 min read
Thread:Agent Rule Compliance
Bingxin Xu and colleagues evaluate whether coding agents—large language models that write robot controllers as programs—can respect a physical safety constraint, pairing each manipulation goal with an obstacle the robot must not touch.
Why this paper
From USC and 2 others · Part of Agent Rule Compliance, now 17 papers
In one line
SafeHarness enables coding agents to prioritize safety, achieving 71.9% success and 87.5% collision avoidance.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§