Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation

Abstract edit

Summary

This paper discusses the Mobile ALOHA system, a low-cost whole-body teleoperation framework designed for bimanual mobile manipulation tasks. It emphasizes the importance of imitating complex mobile manipulation through imitation learning, which combines data collected from its novel teleoperation system with existing static datasets to enhance performance. Mobile ALOHA successfully demonstrates multiple complex tasks, improving success rates significantly when utilizing co-training strategies. The system enables high-performance operation with only a small number of human demonstrations in diverse real-world environments, thus showcasing a new approach to robotic manipulation.

Methods add

This paper employs the following methods:

• Imitation Learning
• Supervised Behavior Cloning
• Co-training

Models Used add

• Mobile ALOHA
• ALOHA
• ACT
• Diffusion Policy
• VINN

Datasets add

The following datasets were used in this research:

• Static ALOHA datasets

Evaluation Metrics add

• Success Rate

Results add

• 95% success on Wipe Wine
• 95% success on Call Elevator
• 85% success on Use Cabinet
• 80% success on Rinse Pan
• 80% success on Push Chairs

Limitations add

The authors identified the following limitations:

• Current footprint of Mobile ALOHA may be too narrow for certain paths
• Fixed height of arms makes lower cabinets difficult to reach
• Robot cannot autonomously improve or explore new knowledge

Technical Requirements edit

• Number of GPUs: 1
• GPU Type: NVIDIA 3070 Ti

Keywords add

Related Papers

Papers with similar methods or datasets:

• Deep Reinforcement Learning from Human Preferences (2017) - 2885 citations
• Multimodal Machine Learning: A Survey and Taxonomy (2017) - 2747 citations
• Deeper Insights into Graph Convolutional Networks for Semi-Supervised Learning (2018) - 2716 citations
• Habitat: A Platform for Embodied AI Research (2019) - 1333 citations
• Machine Learning for Combinatorial Optimization: a Methodological Tour d'Horizon * (2018) - 1306 citations

Papers Using Similar Methods

• Deep Reinforcement Learning from Human Preferences (2017)
• Multimodal Machine Learning: A Survey and Taxonomy (2017)
• Deeper Insights into Graph Convolutional Networks for Semi-Supervised Learning (2018)
• Habitat: A Platform for Embodied AI Research (2019)
• Machine Learning for Combinatorial Optimization: a Methodological Tour d'Horizon * (2018)

External Resources