Solutions
Egocentric video, manipulation demonstrations, tactile data, and robot teleoperation.
Process
Datoric forms a hypothesis from a failure, tests collection methods, then records at scale while checking the method still holds.
01
Datoric starts from a model failure or a capability target and finds the data reasons behind it.
Internal evaluator systems determine hypotheses on the reasons behind the failure and define collection experiments to test them.
What is formed
02
The defined collection experiments are run as side-by-side collections that change one method at a time.
Tests can change cameras, sensors, who demonstrates, the robot, how actions are recorded, objects, rooms, and labels. Video, motion, and tactile methods are tested on their own when the failure points there.
Each test uses the same tasks and the same mix of people, sites, and devices so the only change is the collection method.
What is compared
03
The method that works is then used for the full collection.
New recordings stay checked against the target model and the collection method that already worked. Datoric watches for drift as more hours, sites, and people are added.
Capture stays on the specification that held: specified robot, tasks, sensors, labels, and quality checks.
What continues
Physical AI
UMI Dual-Camera Annotated Manipulation Dataset
500 hours · bimanual demonstrations
Physical AI
Robot Teleoperation Dataset
Custom collection · specified robot
Physical AI
Tactile Glove and Stereo Manipulation Dataset
500 hours · visual and tactile signals
Physical AI
Paired Egocentric RGB and IMU Dataset
20,000 hours · synchronized video and motion
Physical AI
Egocentric Residential Video Dataset
Available · 80 task categories
Physical AI
Multimodal Human Manipulation Dataset
Available · multimodal streams
Physical AI
Industrial First-Person Video Dataset
Available · 1080p+ video
Frequently asked
Egocentric video, manipulation demonstrations, and robot teleoperation. Tasks, equipment, labels, and quality checks are agreed before recording.
A hypothesis is formed from the failure or capability first. Collection methods are tested next. The method that works is then used for the full collection and checked for drift.
Yes. Request samples from the dataset page and specify the tasks and recordings to review.
Teleoperation collections are planned around a specified robot and task list. The number of recordings, sensors, control signals, and file format are agreed before collection.
UMI data uses aligned Zarr episodes. Custom teleoperation collections can use LeRobot, Open X-Embodiment-compatible RLDS/TFRecord, HDF5, ROS bag, or an agreed custom format. Each dataset page lists its recording and delivery formats.
DATORIC