DemoDiffusion: One-Shot Human Imitation using pre-trained Diffusion Policy
Researchers have introduced DemoDiffusion, a novel method for robots to learn manipulation tasks from a single human demonstration without task-specific training. The approach uses kinematic retargeting to convert human hand motion into a robot trajectory, which is then refined by a pre-trained diffusion policy. This ensures the robot's actions are both aligned with the human's movements and plausible within the robot's operational capabilities. In real-world tests across eight diverse tasks, DemoDiffusion achieved an 83.8% success rate, significantly outperforming existing methods. AI
IMPACT Enables robots to learn new manipulation tasks from minimal human input, potentially accelerating robotics deployment.