Ruoshi Wen

“SiMDex uses only ~1.49M mined samples (<5% of the pool) yet improves the overall success rate from 47.7% to 61.1%”

“Our base model follows GR-Dexter [47], a π0-like flow-matching VLA for dexterous manipulation.”

“ByteDance developed GR-Dexter (the base VLA model used in this paper) and the ByteDexter hand”

“Author of both GR-Dexter (the base VLA) and the ByteDexter teleoperation paper [48]. This person is building both the hardware interface and the learning system”

“ByteDance developed GR-Dexter (the base VLA model used in this paper) and the ByteDexter hand... This paper signals ByteDance is building a full dexterous manipulation stack — hardware, VLA model, and now data curation infrastructure.”