Training a robot after it leaves the factory has long meant surrounding it with real hardware. A team at Peking University, working with QiYuan Robotics, says it has found a way to break that dependence.
The result, called HIL-UMI, adds a model brain on top of the UMI data-collection method. UMI, short for Universal Manipulation Interface, is a widely used way to gather manipulation data, but it normally ties the pipeline to physical robots. HIL-UMI gives that collection step a reasoning layer so the post-training loop no longer has to circle a real arm.
The designers are explicit that this is not another data-collection gadget. It is a data flywheel inside post-training: each round of learning feeds the next, so the system compounds instead of restarting. In their measurements, efficiency improved 5.63-fold, and post-training finally stopped revolving around the robotic arm itself.
The bet behind the work is that the next phase of post-training will be won on speed and efficiency, not on how many physical robots a lab can afford to keep running. If the flywheel holds, the bottleneck shifts from hardware to data and reasoning, exactly where Chinese research groups have been investing.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/ai/XgCxObsqu8pc44px.html.