Search papers, labs, and topics across Lattice.
University of Illinois Chicago
6
0
8
UQ rankings for GUI grounding are stable within model classes but falter across different models and interfaces, highlighting the need for tailored calibration in practical applications.
GroundControl reveals that anticipating navigation failures through trajectory-consistent uncertainty can drastically improve the reliability of vision-language agents in real-world applications.
Robots can learn new skills without sacrificing the reliability of previously acquired abilities, achieving a 64% boost in success rates through an innovative wake-sleep consolidation method.
Decomposing uncertainty into aleatoric and epistemic types lets robots recover from errors 21% more effectively than treating all uncertainty the same.
Robots can recover from unexpected mid-episode failures up to 87% faster by adding a small, carefully constrained control signal on top of a pre-trained policy, without any further training.
Transformer-based visual trackers can slash compute by up to 12% without sacrificing accuracy, simply by dynamically adjusting their depth based on uncertainty.
