Simple AI Releases HiFi-UMI-2K Dataset for Robot Manipulation Learning

Simple AI has published a tech report for HiFi-UMI and released HiFi-UMI-2K, a 2,000 hour open dataset for robot manipulation learning, the company announced in a press release.

HiFi-UMI is a portable data production system for collecting human demonstrations used in robot manipulation training. The dataset is released under the Creative Commons Attribution 4.0 license and includes synchronized video from multiple views, bimanual end effector trajectories, gripper states, language annotations, and subtask boundaries.

The report says the system uses head mounted stereo inertial SLAM with measured 3 mm workspace local end effector accuracy. It also uses a shared hardware trigger to align cameras and sensors to below 40 microseconds, and includes automatic trajectory reconstruction and simulation replay validation.

Across three policy backbones and four bimanual tabletop tasks, policies trained further only on HiFi-UMI demonstrations reached success rates comparable to policies trained further on real robot teleoperation data. The reported differences across the three backbones were minus 2.5, plus 3.1, and minus 0.6 percentage points.

Simple AI also reported that pretraining on 4,000 hours of the same corpus reduced offline action prediction error on ten unseen tasks by 41 percent. Human faces in the released recordings are masked.