Search papers, labs, and topics across Lattice.
Affiliation:
5
0
8
12
Neglecting cross-view correspondence can lead to misleading evaluations, with nearly 56% of trajectory pairs showing significant disagreement in agent assessments.
CompCPZ reveals that preserving the full spectrum of user intent in robot manipulation can dramatically improve performance, even in complex, ambiguous scenarios.
Achieving the accuracy of a 12-layer GPT-2 Small with just a 3-layer Gated Recurrent Transformer showcases a groundbreaking efficiency in model design.
Certifying control sequences in the face of model mismatch reveals a trilemma that challenges traditional approaches to safety in sim-to-real transitions.
Set-based training can shrink the reachable-action radius of visuomotor policies without sacrificing performance, outperforming traditional methods by a significant margin.
