Robotics

Scaling Bimanual Household Manipulation from 1,500 hours of Demonstrations to On-Policy Corrections

Jiafeng Xu, Qi Li, Yan Shen, Yiyu Ren, Travis Davies, Shaowen He, Ze Wang, Yifan Yang, Ran Cheng, Hao Dong

Featured September 9, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Training robots to do complex household chores with two arms is hard because good training data is scarce; this paper creates a huge dataset and a smart AI model, XR-2, that learns from both perfect examples and human fixes during mistakes to get much better.

In depth

The paper addresses the challenge of training robust bimanual manipulation policies by introducing a large-scale dataset of 1,500 hours of diverse demonstrations. It then trains XR-2, a powerful vision-language-action (VLA) model, using a multi-stage training paradigm that combines offline expert data with online DAgger correction data to significantly improve performance on deployment-time failures.

Key Takeaways

1
The authors release the PrimeBot Household Manipulation Dataset, a large-scale corpus of 1,500 hours of bimanual demonstrations, combining real-robot teleoperation and UMI data.
2
They train XR-2, a 5B parameter vision-language-action model, demonstrating strong manipulation performance and efficient data utilization.
3
The study shows that combining offline expert demonstrations with online DAgger correction data, particularly with a failure-weighted budget, significantly improves task success rates, especially when expert data alone saturates.

Conceptual Flow

HIGH LEVEL

Methodology: Learning from Examples and Corrections

The paper trains a smart robot brain using a huge collection of how humans do tasks, then lets the robot try and gets humans to fix its mistakes, making it even smarter.

Human Examples

Robot Tries

Human Fixes

Learn and Improve

Smart Robot Brain

Results: Significant Performance Gains

By using lots of examples and then fixing mistakes, the robot's ability to fold clothes jumped from okay to almost perfect.

Robot's First Try

More Examples

Human Corrections

Big Improvement

Robot Does Task Well

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.