Cross Embodiment

Cross-embodiment research aims to enable robots to learn and perform tasks across diverse physical forms (embodiments), transferring skills learned on one robot to another with different capabilities. Current research focuses on developing algorithms and model architectures, such as transformers and diffusion models, that learn robust representations of actions and environments, facilitating zero-shot transfer to unseen robots. This work is significant because it promises to drastically reduce the cost and time required to train robots for new tasks, paving the way for more adaptable and versatile robotic systems in various applications.