Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation

Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation. By Hyunwoong Bae and four others, from Seoul National University. Published on the twenty-sixth of October.

A robot moving through a home should not treat every obstacle alike. Safe household navigation means regulating clearance and speed according to what an obstacle is and how close it is, and doing that without being needlessly conservative. Existing evaluations, the authors argue, do not test whether semantic safety distinctions actually show up in continuous motion.

The measuring instrument is the Semantic Sensitivity Index, or SSI. It asks whether clearance, together with speed and smoothness as they vary with distance, follows the caution ordering of the obstacles. A robot that recognizes one obstacle as more sensitive than another should move differently around it, and the index checks whether it does.

The benchmark is OSMA-Bench, short for Obstacle-Conditioned Safe Motion Adaptation Benchmark. It is built on scenario sets that are spatially matched and differ only in the type of obstacle. Holding the layout fixed means any change in motion can be attributed to what the obstacle is. The same room, the same path, the same distances: only the identity of the thing in the way is different. If a policy treats two such scenarios identically, its motion is not reflecting the safety distinction between their obstacles.

Policies of two kinds are evaluated: VLM-based policies, and vision-language-action or VLA policies. Neither consistently adapts its motion to obstacle safety, even when given explicit safety prompts. Telling a policy to be careful, in words, is not enough to make its movement careful. The analysis then locates where the failure happens. The policies can correctly tell obstacles apart by safety relevance. What they fail to do is carry that distinction into the motion parameters or actions they produce. For VLM-based policies, which recognize the obstacle and then generate a navigation program, inserting an explicit parameter-inference step between those two stages improves adaptation.

The gap exposed here is in semantic-to-motion grounding. Knowing that one obstacle calls for more caution is not the same as moving more cautiously around it, and these policies manage the first without the second. The authors highlight the need to translate semantic caution into motion that is appropriately safe.