Zeyi Liu and Shuran Song have presented BATCH, a framework that explicitly aligns human and robot behaviour so that robots can learn manipulation from human demonstrations, in a paper posted on arXiv.
BATCH learns temporal representations with a cycle-consistency objective to find correspondences between human and robot demonstrations without manual labels. It uses them in an autoregressive video-action diffusion model, both for co-training and for transferring human style to the robot.
On real-world manipulation tasks, it generalised better to unseen task setups than training on unaligned data or earlier cross-embodiment methods, the authors report.
Background
Zeyi Liu
Robotics researcher
Zeyi Liu is first author of the BATCH paper on learning robot skills from human demonstrations.
