University of Maryland

“By physically removing transformer blocks and then fine-tuning the remaining model (a protocol they call Drop-Then-Recovery, or DTR), the authors found that language capacity is highly redundant.”

“"language backbones are highly redundant for standard robotic manipulation tasks, whereas vision and action pathways are substantially less tolerant to removal."”

“"The model runs on an NVIDIA Jetson Thor" (Section 5).”

“"Training data are collected via teleoperation with Meta Quest 3 at 10 Hz" (Section 5).”

“"a wrist-mounted Intel RealSense D435 camera" (Section 5).”

“"Ramana Rao Kompella 2, Gaowen Liu 2... 2 Cisco Research" (Title page).”