Search papers, labs, and topics across Lattice.
Affiliation:
2
0
2
Standard robot policies collapse on non-Markovian tasks requiring temporal tracking and recall, but pairing Mamba-2 structured memory with flow matching boosts real-world success from 8.3% to 66.7% while revealing that counting and spatial memory demand fundamentally different recurrent update rules.
Decoupling modality processing in VLA models leads to a staggering 95.2% success rate in complex manipulation tasks, far surpassing traditional synchronous approaches.
