Long-Horizon Language-Conditioned Imitation Learning for Robotic Manipulation

摘要

Language-controlled policies enable robots to follow human language instructions and execute complex tasks. While language-conditioned imitation learning has proven effective in teaching robots to perform tasks guided by language instructions, it faces multiple challenges due to the multimodal nature of human demonstrations and limited training data. The variability in demonstrations can complicate policy learning, as the same instruction may correspond to diverse actions. To mitigate these issues, we propose an end-to-end transformer-based policy, predicting categorical distributions over a discretized action space. By discretizing the action space and employing autoregressive sampling, our model efficiently handles the exponential growth of high-dimensional discrete action spaces, allowing it to learn complex action distributions effectively. In addition, we apply data augmentation techniques to reuse existing data more effectively and implement an action disturbance strategy to enhance the model's generalization capabilities. Furthermore, we employ a cotraining strategy to leverage data that lacks language annotations. The effectiveness of our approach is demonstrated through simulation and real-world experiments on a robot manipulator in a long-horizon, language-conditioned setting, including multiple environments and zero-shot transferring to real-world settings.