Original title: Scaling Bimanual Household Manipulation from 1,500 hours of Demonstrations to On-Policy Corrections
Jiafeng Xu, Qi Li, Yan Shen, Yiyu Ren +6 · September 2026
Teaching robots to use two arms for household chores is difficult because good training data is rare. To fix this, researchers collected 1,500 hours of video showing humans performing everyday two-handed tasks. They used this data to train a robot control system called XR-2. The system learns through a multi stage training setup that translates sight and words into physical motions efficiently. In tests, the robot showed strong manipulation skills. The team also discovered that the robot gets steadily better under two conditions: when fed more demonstration data, and when humans step in during real time to correct its mistakes. These consistent gains show that collecting more practice data directly improves how well the robot handles two-handed tasks.
You want to learn a complex two-handed task, like folding laundry. At first, you watch someone else do it, but you keep making mistakes because you don't see every detail. In this paper, robots learn by watching 1,500 hours of humans doing chores, and then humans correct their errors to help them get better.
For people building robots to assist in homes, having more demonstration data means robots can learn a wider range of tasks more reliably. This could lead to robots that help with chores like folding laundry or preparing meals.
The abstract does not explicitly state limitations, but the work is focused on bimanual household manipulation and the scaling trends are studied within specific data ranges probed in the experiments.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Scaling Bimanual Household Manipulation from 1,500 hours of Demonstrations to On-Policy Corrections — Jiafeng Xu, Qi Li, Yan Shen et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.