Track synthesis, materials and catalysis as it is published.
Search papers, labs, and topics across Lattice.
Affiliation:
2
0
0
A simple test-time trigger and latent world model can boost VLA policy success by 15% without retraining, by intelligently selecting the best action from stochastic outputs.
Simply handing VLMs a 360° field of view fails to improve embodied navigation unless paired with dynamic multi-step action planning and explicit geometric representation, which together yield an 11.9% jump over SOTA.