Generalizing manipulation policies across different robot bodies is hard because standard policies tie task-relevant interactions to a specific robot's actions. In a new paper, researchers introduce KITE (Kinematic Interaction Transfer across Embodiments), a framework for zero-shot cross-embodiment manipulation. The system decouples the problem into two parts: predicting an embodiment-agnostic interaction intent, based on contact patterns, and decoding that intent into robot-specific motions.
The interaction policy is trained only on demonstrations collected from a source robot, then shared across morphologically different target robots. A kinematic decoder handles each target by learning from synthetic data generated from that robot's kinematic model. According to the paper, this means a new robot can adopt a learned skill given only its kinematic model, with no need to recollect task-specific demonstration data.
The authors evaluate KITE on four manipulation tasks in both simulation and the real world, covering transfers between parallel grippers, dexterous hands, and composite embodiments. They report that KITE achieves successful zero-shot transfer to structurally different target embodiments and consistently outperforms state-of-the-art baselines.