Search papers, labs, and topics across Lattice.
Affiliation:
3
0
4
Visual token pruning often degrades robotic control because it ignores language conditioning; steering token cache reuse via text-vision entropy differentials cuts VLA compute by 2.45x while boosting task success by up to 14.5%.
Grasping performance can be dramatically improved with a single view adjustment that leverages affordance fields, bypassing the need for complex scene reconstruction.
VLAConf offers a computationally cheap, architecture-agnostic way to estimate task success confidence for robot manipulation by turning frozen VLA representations into anomaly scores.
