Matteo Matteucci

State-of-the-art VLA models exhibit significant behavioural divergence under perturbation, even when Task Success Rate remains unchanged, revealing a critical blind spot in current robustness evaluations.

Pedestrians alter their velocity and perception when looking directly at an approaching robot, and incorporating these egocentric awareness cues alongside patch-partitioned occupancy maps in a selective state-space backbone establishes new state-of-the-art accuracy across five navigation benchmarks.