Simple AI's robot-free dataset matches teleoperation at more than ten times the demonstrations

Xiaofei Li, founder and chief executive of Simple AI, released a 2,000-hour robotics dataset on August 20 collected without a target robot or robot teleoperation anywhere in the loop ^[1]. The accompanying technical report, first posted on July 28, claims approximate performance parity with conventional robot teleoperation across three policy architectures and four tabletop task groups ^[2], and the price of that parity is stated in the same materials: more than ten times as many demonstrations ^[3].

That ratio is the actual offer. Teleoperation data is expensive because every recorded action already encodes a specific machine's geometry, sensors and limits, which also ties each hour of collection to an arm, a control rig and an operator and makes expansion across homes, hotels and warehouses slow and capital-intensive ^[4]. Simple AI's HiFi-UMI system instead records people performing two-handed tasks with portable sensor-equipped grippers, then converts the demonstrations into trajectories for real arms ^[5]. On the reported evidence, each robot-free demonstration is worth under a tenth of a teleoperated one at the margin ^[6]. Whether that is a good trade depends entirely on whether your bottleneck is hardware-hours or storage and processing.

The approach descends from the Universal Manipulation Interface, a robot-free human-demonstration system introduced in a 2024 paper by researchers affiliated with Stanford University, Columbia University and Toyota Research Institute ^[7]. Simple AI's work targets the leftover fidelity problems: trajectory accuracy, synchronization, camera coverage and the relative position of the two grippers ^[8]. The build puts a stereo camera and inertial sensors on the operator's head and two non-parallel fisheye cameras on each hand module, with a shared hardware trigger aligning them ^[9]. Simple AI reports roughly 3 millimeters of local end-effector accuracy, sensor synchronization under 40 microseconds and about 200 degrees of visual coverage per hand ^[10].

The unglamorous part is the rejection step. Each trajectory is reconstructed and replayed in simulation, with reported pass rates of about 98% at both stages ^[11], which compounds to roughly 96% surviving if the gates run in sequence ^[12]. Cheap collection stops being cheap the moment engineers have to inspect episodes by hand.

HiFi-UMI-2K holds 2,000 hours and more than 482,000 episodes across over 110 scenes, per its dataset card, with synchronized multi-view video, bimanual trajectories, gripper states, language annotations, subtask boundaries and quality-control metadata, under CC BY 4.0 ^[13]. That works out to about 15 seconds per episode ^[14]. Simple AI says the release is less than a tenth of a processed corpus of more than 20,000 hours and 4.32 million episodes across over 480 scenes ^[15]; by hours the open share is 10%, by episode count about 11% ^[16].

On the numbers themselves: against policies post-trained on teleoperation data, the three HiFi-UMI-only backbones produced aggregate success-rate differences of -2.5, +3.1 and -0.6 percentage points ^[17]. Those three values average exactly zero across a 5.6-point spread ^[18], which is a defensible reading of parity and also a small sample. On a precision-insertion task the strongest HiFi-UMI-only policy hit 85% success, and the report notes the teleoperation baseline had an apparent environmental advantage there ^[19].

Watch whether outside teams reproduce parity using only the released 2,000 hours, since the internal result was presumably tuned against the full corpus ^[15][2]. Watch the demonstration ratio: if it drops below ten with better fidelity, the argument gets stronger; if it climbs, the storage and compute bill eats the hardware savings ^[3][6].