Democratizing Data-Efficient Robot Learning with Robot-Free Demonstrations

Collecting high-quality robot demonstrations remains a major practical bottleneck for robot learning. Teleoperation requires continuous access to the target robot and can be substantially slower than natural human manipulation. To help democratize data-efficient robot learning we introduce an affordable, self-buildable, co-designed data glove and matched robot hand with $8$ active degrees of freedom, that costs only $300$ dollars to build. The system enables robot-free collection of synchronized wrist observations, wrist motion, and finger configurations during natural human manipulation. Demonstrations require only lightweight post-processing. Visible human skin is normalized to match the glove, and demonstrations are temporally aligned with robot execution, which is typically slower than natural human manipulation. Unlike some target-specific UMI pipelines, which replay every demonstration on the physical robot hand, our method requires replay. For example, collecting one demonstration takes approximately half the time compared to DexUMI, a similar collection-and-replay system. We evaluate how downstream policy performance scales with the number of robot-free demonstrations across $2$ dexterous manipulation tasks. Using ACT and GR00T, we compare our collection method against direct teleoperation. For example, our method achieves $80\%$ success on an Egg Carton task with $150$ robot-free demonstrations. In total collection time, this is comparable to approximately $30$ teleoperated demonstrations or $300$ DexUMI demonstrations. These results show that inexpensive, rapidly collected human demonstrations can substantially reduce the robot access and collection time required for effective real-world manipulation learning. Our full hardware specifications and open-source software will be made publicly available.