Cross-Embodiment Fine-Tune

Key Insight

Cross-embodiment learning allows a robotic policy to generalize across different robot geometries, actuators, and platforms by pretraining on massive, heterogeneous datasets like Open X-Embodiment. Fine-tuning these large-scale models on a specific target robot requires far fewer demonstrations than training from scratch, because the model's vision encoder and spatial representations are already optimized. By learning a shared action-representation mapping, the robot leverages general physical concepts learned from other platforms to solve its specific task with high sample efficiency.