Wire · founder news, decoded · technology
Humanoid Robots Can't Handle Human Touch: New Benchmark Exposes VLA Safety Gap
ThorArena, a new benchmark from TU Munich, exposes critical safety gaps in vision-language-action (VLA) models used in humanoid robots—all leading models fail to reliably handle direct physical human contact, a scenario absent from existing ISO and NIST standards but increasingly common in real-world deployment.
This Wire brief sits within Fusion42's coverage of AI Agents and Robotics. Wire is Fusion42's founder-focused intelligence feed: each story is connected to the funds and startups it names — every one with a live profile on Raise or Scout — so founders can follow the capital and the momentum behind the headline rather than just the headline itself. Wire analysis is one of the live surfaces Arthur, Fusion42's AI co-founder, reasons over.
The Wire takeaway
Founders building humanoid robot control systems, VLA training pipelines, or safety certification tools now have concrete evidence that existing benchmarks miss critical failure modes; ThorArena creates both a regulatory/standards opening and a product differentiation vector for teams solving force/contact-aware robot policies.
Read the full story at techtimes.com →
Topics: AI Agents · Robotics · humanoid-robots · vla-models · safety-benchmarking · physical-human-robot-interaction · robot-certification · standards-gap