Imitation Learning Overview
DISCOVERSE provides a complete imitation learning workflow, supporting end-to-end training from data collection to policy deployment. This chapter introduces four mainstream imitation learning algorithms integrated in the framework.
đ¯ Supported Algorithmsâ
Currently, DISCOVERSE supports the following four imitation learning algorithms:
1. ACT (Action Chunking with Transformers)â
- Data format: HDF5
- Use case: Complex manipulation tasks requiring long sequence planning
2. DP (Diffusion Policy)â
- Data format: Zarr
- Use case: Multimodal action distributions, complex manipulation skills
3. RDT (Robotics Diffusion Transformer)â
- Data format: HDF5
- Use case: Multi-task learning, general robot skills
4. OpenPI (Open-source Policy Interface)â
- Data format: HDF5
- Use case: Rapid prototyping, few-shot learning
đ Workflowâ
The complete imitation learning workflow includes the following steps:
1. Data Generationâ
Automatically generate demonstration data, 100x more efficient than the real world
2. Data Format Conversionâ
Convert to the required format for each algorithm:
- ACT/RDT/OpenPI: Raw data īŋŊ?HDF5
- DP: Raw data īŋŊ?Zarr