The paper "JUICER: Data-Efficient Imitation Learning for Robotic Assembly" addresses the challenge of learning precise, long-horizon manipulation tasks from limited human demonstrations. The authors propose a pipeline called JUICER, which combines expressive policy architectures and techniques for dataset expansion and simulation-based data augmentation. The pipeline is designed to improve imitation learning performance with a small number of demonstrations, particularly for tasks requiring precise grasping, reorienting, and inserting multiple parts over long horizons. The key components of JUICER include: 1. **Data Collection and Annotation**: Collecting a small number of demonstrations (~50) and annotating "bottleneck" states requiring high precision. 2. **Trajectory Augmentation**: Expanding the dataset by synthetically creating corrective actions around these bottleneck states through disassembly and reassembly sequences. 3. **Policy Design**: Using diffusion policy architectures to model action distributions conditioned on observations, with a focus on long-term action consistency. 4. **Dataset Expansion with "Collect-and-Infer"**: Incorporating successful rollouts from model evaluations back into the training set to further expand the dataset. 5. **Multitask Learning**: Training models on a mix of demonstrations from different tasks to leverage shared skills and improve performance. The effectiveness of JUICER is demonstrated on four furniture assembly tasks in simulation, achieving high success rates even with a modest number of demonstrations. The paper also discusses related work, limitations, and future directions, highlighting the potential for applying the JUICER pipeline in real-world settings.The paper "JUICER: Data-Efficient Imitation Learning for Robotic Assembly" addresses the challenge of learning precise, long-horizon manipulation tasks from limited human demonstrations. The authors propose a pipeline called JUICER, which combines expressive policy architectures and techniques for dataset expansion and simulation-based data augmentation. The pipeline is designed to improve imitation learning performance with a small number of demonstrations, particularly for tasks requiring precise grasping, reorienting, and inserting multiple parts over long horizons. The key components of JUICER include: 1. **Data Collection and Annotation**: Collecting a small number of demonstrations (~50) and annotating "bottleneck" states requiring high precision. 2. **Trajectory Augmentation**: Expanding the dataset by synthetically creating corrective actions around these bottleneck states through disassembly and reassembly sequences. 3. **Policy Design**: Using diffusion policy architectures to model action distributions conditioned on observations, with a focus on long-term action consistency. 4. **Dataset Expansion with "Collect-and-Infer"**: Incorporating successful rollouts from model evaluations back into the training set to further expand the dataset. 5. **Multitask Learning**: Training models on a mix of demonstrations from different tasks to leverage shared skills and improve performance. The effectiveness of JUICER is demonstrated on four furniture assembly tasks in simulation, achieving high success rates even with a modest number of demonstrations. The paper also discusses related work, limitations, and future directions, highlighting the potential for applying the JUICER pipeline in real-world settings.