SafeHarness: Coding Agents That Grasp Without Collision
A preprint reports that coding agents usually hit obstacles while grasping, and that a two-part harness lifts collision avoidance to 87.5 percent.
In short
No — the preprint finds that coding agents driving a robot arm collide with the obstacle in most cases, and only its proposed SafeHarness layer brings them to 87.5% collision avoidance at a 71.9% task success rate.
At a glance
- SafeHarness reports 71.9% task success and 87.5% collision avoidance.
- The stated margin over the previous state of the art is 6.5% and 27.0%.
- Against the agent with no harness that works out to 2.3x and 1.5x respectively.
- Two parts: route planning with bounding boxes, waypoint checks and replanning, plus obstacle-aware contact execution.
- Preprint dated September 17, 2026, arXiv 2609.20822, filed under cs.RO, cs.AI, cs.CL and cs.CV; not peer reviewed.
A coding agent writes the robot controller as a program, which is how these systems operate hardware they were never trained on. Bingxin Xu, Yuzhang Shang, Zhen Dong and Emilio Ferrara ask the question that this line of work has skipped: is the resulting behavior safe? Their abstract answers bluntly — the agent goes after the goal and hits the obstacle in most cases.
A planning failure, not a hardware one
The authors trace the collisions to two moments: choosing the route, and executing the contact. The agent treats finishing the task as the whole job, so the obstacle never enters the generated program as a constraint. That makes the crash a property of the code the model wrote, not of the gripper.
What the harness adds
SafeHarness wraps the model in two layers. The first plans the route up front, representing obstacles as bounding boxes, verifying each waypoint, and replanning when a waypoint passes through something solid. The second governs the grasp itself, picking a contact position that already routes around the obstacle.
Both interventions sit at the level where the failure originates. Nothing about the underlying language model changes, and no robot-specific training is involved.
Reported results
The abstract puts SafeHarness at 71.9% task success and 87.5% collision avoidance, ahead of the previous state of the art by 6.5% and 27.0%. Measured against the bare agent with no harness at all, that is 2.3x on success and 1.5x on collision avoidance.
What could not be checked
The public abstract names no simulator or benchmark, gives no task count, and does not say which language models were tested, so none of that could be verified here. It also leaves open whether 6.5 and 27.0 are percentage points or relative gains. This is a preprint from September 17, 2026 with no peer review yet; the record sits on arXiv.
FAQ
What is SafeHarness?
A harness layer for coding agents that plans robot routes around obstacles and chooses a grasp contact point that avoids them. It leaves the underlying language model untouched.
Why do coding agents crash robot arms into obstacles?
They optimize for completing the task, and the obstacle never shows up as a constraint in the program they generate. The preprint locates the failure in route selection and contact execution.
How much does SafeHarness improve over prior methods?
The abstract reports 71.9% task success and 87.5% collision avoidance, ahead of the previous state of the art by 6.5% and 27.0%.
