TCP4D: A Unified Tool-Center-Point Tracking and Generation Model for Robot Manipulation
TCP4D turns monocular robot video into motion and gripper-state annotations without recorded source-robot states.
A shared 4D scene representation recovers tool-center-point poses and supports future trajectory prediction guided by robot observations and language instructions. Its augmentation pipeline retargets recovered motion to other robot embodiments and synthesizes corresponding observations. The authors report tracking errors of 7.60 mm on DROID and 16.07 mm on RoboLab, plus adaptation to a new embodiment with ten demonstrations. The generator reaches 69.19% mean success on RoboTwin; combining source and augmented demonstrations improves target-robot success by 41.5% over augmented-only training.
HF Daily Papers' note
A shared 4D scene representation recovers tool-center-point poses and supports future trajectory prediction guided by robot observations and language instructions. Its augmentation pipeline retargets recovered motion to other robot embodiments and synthesizes corresponding observations. The authors report tracking errors of 7.60 mm on DROID and 16.07 mm on RoboLab, plus adaptation to a new embodiment with ten demonstrations. The generator reaches 69.19% mean success on RoboTwin; combining source and augmented demonstrations improves target-robot success by 41.5% over augmented-only training.
HF Daily Papers' note
score 5