CosmoH2G: A Hand-to-Gripper Transfer Dataset and Baseline Method for Object Manipulation with Complex Spatial Movements
The paper’s core claim is that hand-to-gripper transfer needs paired data with harder 3D motion, not just better imitation on planar tasks.
CosmoH2G provides 6,189 paired demonstration episodes over 1,254 objects, collected with a handheld gripper to mimic fine-grained hand motion. The authors say existing approaches break down on rotations, flips, and other complex spatial trajectories. Their baseline predicts sparse initial and terminal gripper keyframes first, then fills in the continuous action sequence while post-optimizing translation to reduce drift. They report stronger results than traditional baselines in simulation and real-robot tests. HF Daily Papers' note
CosmoH2G provides 6,189 paired demonstration episodes over 1,254 objects, collected with a handheld gripper to mimic fine-grained hand motion. The authors say existing approaches break down on rotations, flips, and other complex spatial trajectories. Their baseline predicts sparse initial and terminal gripper keyframes first, then fills in the continuous action sequence while post-optimizing translation to reduce drift. They report stronger results than traditional baselines in simulation and real-robot tests. HF Daily Papers' note
score 4