Progress in robot manipulation learning is increasingly constrained by data. Real-robot teleoperation yields accurate, directly trainable trajectories but is difficult to scale: every hour of data requires the target robot, a teleoperation rig, and a skilled operator. Robot-free handheld demonstrations are cheaper and easier to scale, but have primarily been used for pre-training. Task-specific post-training, the stage that grounds a policy for real-robot deployment, has typically continued to rely on a smaller amount of real-robot teleoperation data as an anchor.<\/p>\n The report examines whether raising the fidelity of robot-free demonstration data, rather than shrinking the real-robot fraction, can remove that anchor for target-task post-training.<\/p>\n HiFi-UMI is a portable data-production system co-designed for four fidelity properties. Pose accuracy comes from head-mounted offline stereo-inertial SLAM, which the report measures at 3mm workspace-local end-effector accuracy. Cross-sensor timing is aligned to below 40 microseconds through a shared hardware trigger across all cameras and sensors. Inter-gripper relative pose is measured natively rather than reconstructed. Per-hand sensing covers approximately 200 degrees of field of view through two non-parallel wide-angle fisheye cameras. Every captured demonstration passes through automatic trajectory reconstruction and simulation replay validation, each gate with an approximately 98% pass rate.<\/p>\n The report evaluates the approach across three policy backbones spanning the vision-language-action and world-action-model families, and four bimanual tabletop tasks. Policies post-trained solely on HiFi-UMI demonstrations reached success rates comparable to policies post-trained on in-domain real-robot teleoperation data, with reported differences of \u22122.5, +3.1, and \u22120.6 percentage points across the three backbones. On a precision insertion task, the strongest HiFi-UMI-only policy reached 85% success under conditions where the teleoperation baseline had the advantage of being collected in the evaluation scene. Separately, pre-training on 4,000 hours of the same corpus reduced offline action prediction error on ten unseen tasks by 41%, and increased real-robot success on one of the evaluated backbones by 18.1 percentage points at matched post-training data.<\/p>\n “We wanted to test whether fidelity, rather than scale alone, is what unlocks robot-free data for deployment-oriented training,” said Xiaofei Li, founder of Simple AI. “The report shows what this can look like within a specific set of tasks and models. By open-sourcing HiFi-UMI-2K, we hope to give the wider research community a shared, high-fidelity resource for continuing to study this question.”<\/p>\n The report characterizes these findings as approximate aggregate parity within the tested models, tasks, and experimental conditions. The deployment robot uses the same gripper and wrist-camera configuration as the capture setup, with the main embodiment difference being robot arm kinematics. The comparison is not sample-matched, with 3,200 HiFi-UMI trajectories per task set against approximately 300 teleoperation trajectories, and reflects a comparison between practical data-production pipelines rather than a claim of per-trajectory equivalence. The report does not generalize the result to all robot learning settings, and does not conclude that real-robot data is no longer required in the broader field.<\/p>\n HiFi-UMI-2K is distributed in a training-ready format with synchronized multi-view video, bimanual end-effector trajectories, gripper states, language annotations, and subtask boundaries. Human faces in the recordings are masked before release. The paper reached No. 1 on Hugging Face Daily Papers on July 29.<\/p>\n HiFi-UMI is one component of Simple AI’s work across foundation models, high-fidelity data, robotic systems, and real-world deployment. The company welcomes conversations with research groups and industry partners interested in high-fidelity data for robot learning.<\/p>\n About Simple AI<\/b> Resources<\/b><\/p>\n Tech Report: <\/b>arxiv.org\/abs\/2607.25895\u00a0<\/p>\n Dataset:<\/b> huggingface.co\/datasets\/simple-world-lab\/HiFi-UMI-2K\u00a0<\/p>\n Project page:<\/b>\u00a0cloud.simpleai.tech\/simple-world-lab\/hifi-umi\/\u00a0<\/p>\n Media Contact<\/b><\/p>\n Grant Xin<\/p>\n Simple AI<\/p>\n![\"A](\"https:\/\/mmx.prnasia.com\/media\/MS1973087\/20260820084529EDT_image_1.jpg?id=OA2900133&p=medium600\" "\\"A")
<\/a>
A portable, high-fidelity robot-free data-production system for robot manipulation learning. Across three policy backbones evaluated in the report, policies post-trained solely on HiFi-UMI data reached success rates comparable to policies post-trained on in-domain real-robot teleoperation data.<\/span><\/p>\n<\/div>\n
Simple AI is an embodied AI company developing general-purpose embodied intelligence systems for human living spaces. Its work integrates foundation models, high-fidelity data, robotic systems, and real-world deployment across the full embodied AI stack.<\/p>\n