What Data Robots Need: From Co-Designed Handheld Capture to Fleet-Scale Flywheels (RSS 2026)

Invited talk at "It's the Demos: A Deep Look at the Role of Demonstration Quality in Imitation-Based Robot Manipulation," a workshop at Robotics: Science and Systems (RSS) 2026, Sydney, Australia, July 13, 2026. Imitation learning is only as good as the demonstrations behind it. This talk covers what we've learned building robot data collection at scale: co-designing handheld capture hardware (hardened UMI-style capture gloves and a Spot gripper glove) together with the policies they feed, how our data collects compare against standard UMI, why loss curves are (mostly) meaningless for policy quality, bridging handheld and teleoperated supervision to fix failure modes, and maintaining demonstration quality in a 100-robot teleoperation pipeline toward a fleet-scale data flywheel. Workshop: https://its-the-demos.github.io/ Chapters: 0:00 Title & intro 0:45 Background on UMI 1:49 Loss curves are (mostly) meaningless 2:32 Hardened UMI capture glove 3:15 Spot gripper capture glove 4:08 How do we compare our data collects? 6:28 Policy results in unseen environments 10:22 Bridging handheld and teleoperated supervision 11:08 Modified diffusion policy architecture 12:23 Codesigning grippers 14:18 Maintaining demonstration quality in a 100-robot teleoperation pipeline 15:07 Data factory 17:31 Contributions 17:42 The physically grounded Turing test 18:09 Summary and next steps 19:08 Closing demo & Q&A