Search papers, labs, and topics across Lattice.
4
0
7
Pre-imagined milestones can boost VLA recovery success rates by over 30% without requiring low-level action adjustments.
Forget scaling alone: co-designing sensing and inference in a bio-inspired architecture yields a 34% accuracy boost for embodied AI.
Even when data distributions shift, in-distribution and out-of-distribution samples remain surprisingly separable: DART dynamically tracks this "discriminative axis" to boost OOD detection by 15% AUROC under heavy corruption.
Achieve state-of-the-art multivariate time series imputation by binding CNN channels to Transformer heads, improving robustness to missing data by selectively down-weighting unreliable temporal patterns.
