Humanoid safety filters fail under ordinary goal-insertion attacks
Original title: Not an Accident: Exploiting Safety Filter Inadequacy in Humanoid Robots
SIREN found attacks in every vulnerable SPARK scenario and caused a real humanoid to collide during tasks.
A paper argues collision-avoidance guarantees for humanoid robots hold only when safety-filter parameters match the task and environment, and no general method sets those parameters. Filters shipped with robots can therefore fail after deployment. Attackers can insert goals through normal user interaction without extra privileges. The authors' tool SIREN found such attacks in every vulnerable SPARK scenario and generated a collision on a physical humanoid.
Unconfirmed: The demonstrations are experimental; no commercial model, deployment, or CVE is named.
- SIREN discovered three times as many attacks as a baseline.
- It identified attacks in all vulnerable SPARK scenarios.
- A real humanoid collided while executing tasks.
- No extra privileges beyond normal user interaction were required.
Impact: Humanoid collides with nearby people or objects
| Product | Affected |
|---|---|
| humanoid robot safety filters | — |
From the original article
1 min read at arxiv.orgAs humanoid robots move from research laboratories toward deployment in human-centered environments such as homes, hospitals, and warehouses, safe operation around nearby people and objects becomes a fundamental requirement. Safety filters have emerged as a promising technique to address this challenge by providing mathematical guarantees that generated control actions do not result in collisions.
By Tiffany Bao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.