Skip to content

Back to the catalog

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

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

This arXiv paper studies coding agents that write robot controller programs for manipulation tasks under a safety constraint (avoid touching a specified obstacle), finding they collide in most cases despite reasoning about the obstacle and being explicitly instructed not to touch it. The authors trace the failure to planning: the model lacks a notion of a clearing route and cannot replan when a route becomes infeasible, and treats contact execution as unbounded by the constraint. They propose SafeHarness, an obstacle-aware harness with route planning (bounding boxes, candidate waypoint routes, verification and replanning) and obstacle-aware contact execution, which raises task success to 71.9% and collision avoidance to 87.5%, beating prior SOTA by 6.5% and 27.0% respectively.

TL;DR A harness (SafeHarness) that makes LLM coding agents plan obstacle-aware routes and contact points, cutting collisions in robot manipulation tasks by prioritizing safety constraints the model otherwise ignores.
Open original ↗
Document

What it adds beyond the summary

The catalog summary already carries the framing, the diagnosed failure, the two harness components, and the headline numbers (71.9% task success, 87.5% collision avoidance, +6.5% / +27.0% over prior SOTA). The one thing it omits is the ablation against the naked agent: SafeHarness's results are 2.3× the task success and 1.5× the collision avoidance of the same agent without the harnesses. That implies the base coding agent lands near ~31% success and ~58% collision avoidance on this constrained setting — the gap the harness closes, and the reason the paper exists rather than a tuning note.

Evidence, as stated

  • Task success: 71.9%; collision avoidance: 87.5%.
  • Margins over previous SOTA: +6.5% (success) and +27.0% (collision avoidance) — note the safety gain is far larger than the capability gain.
  • Ablation: 2.3× (success) and 1.5× (collision avoidance) versus the identical agent with no harnesses.

The page is abstract-only, so no dataset name, task count, model identity, compute, or per-task breakdown is given — the absolute baseline figures above are inferred from the multipliers, not stated directly.

Worth knowing for adoption decisions

The paper's decomposition is the reusable idea, not just the numbers: it splits manipulation into a route phase and a contact-rich moment and attacks each separately. Route planning grounds objects as bounding boxes, draws candidate waypoint routes, and runs a plan → verify → replan loop before execution. Contact execution is treated as a distinct problem: it selects a contact position so the contact itself avoids the obstacle. If you are building safety-constrained coding agents, the actionable claim is that instruction and perception were not the fault — the model saw and was told about the obstacle — so prompt fixes will not help; the constraint has to be made a first-class planning object.

Limits

Abstract only, so no code, weights, benchmark name, or task suite is confirmed on this page — check the PDF before assuming reproducibility. Even with the harness, ~12.5% of runs still collide and ~28% fail the task, so this is a mitigation, not a guarantee, for any real deployment.

Techniques named here

  • SafeHarness
  • Obstacle-aware route planning
  • Obstacle-aware contact execution

Related in research