Mingyuan Zhang

Robot manipulation policies can achieve state-of-the-art performance and 17x faster inference by freezing the VLM and using a lightweight actor to query a fixed task memory.

GlanceWAM achieves 72.2% success on RoboCasa while executing actions 24 times faster than traditional synchronous models, reshaping the landscape of real-time robot learning.

Every exactly calibrated convex surrogate for the Jaccard measure demands exponentially more prediction dimensions than previously understood.