Egocentric human demonstrations
Continuous first-person demonstrations capturing hands, tools, objects and surrounding workspace context.
RGBIMUAudio where configuredTimestampsCalibration metadata
DATRAAI / REAL-WORLD DATA INFRASTRUCTURE
Build the real-world data distribution your model is missing.
DatraAI designs and operates custom data programs for robotics, world-model and embodied-AI teams—capturing egocentric, stereo, UMI-style and teleoperation data, then delivering raw, synchronized or training-ready episodes.
Define the tasks, environments, sensors, embodiment, volume and acceptance criteria. We design the pilot around them.
Custom data programs for physical AI01 / 04
Explore multimodal dataBacked by
Founders, Inc.
The Residency
Inception programProgram and investor affiliations shown for company context; they are not presented as customer relationships.
Robotics foundation modelsWorld & action modelsVLA post-trainingManipulation & teleopMultimodal pretraining
01 / Multimodal Data
Multimodal data built around your target distribution.
You define the tasks, environments, modalities, geography, volume, output format and acceptance criteria. DatraAI designs the capture configuration, operates the data program, validates each session and delivers the required package.
Stereo spatial capture
Synchronized dual-view global-shutter capture for geometry-aware human activity and manipulation datasets. The stereo signal supports downstream geometry and disparity processing.
Left RGBRight RGBHigh-rate IMUAudio where configuredCamera intrinsicsCamera extrinsicsSynchronized timestamps
Portable manipulation capture
UMI-style and gripper-based collection for contact-rich manipulation, including device pose, gripper state, visual context and tactile information where configured.
RGBGripper poseGripper openingIMU / VIOVisuotactile forceTimestamps
Robot teleoperation
Robot observations, state and actions captured during human-controlled execution, scoped to the customer’s target embodiment and task set.
Camera observationsRobot stateActionsEnd-effector stateTask metadata
Available delivery levels
Raw sensor data
Original video, IMU, audio, pose, tactile and robot streams with associated metadata.
Synchronized sessions
Time-aligned streams, calibration, session manifests and quality reports.
Training-ready episodes
Structured episodes packaged through Actuate into LeRobot, RLDS, canonical or agreed customer formats.
EnvironmentsPrograms can be scoped for commercial kitchens, hospitality, retail, warehousing, manufacturing, agriculture, home, office and laboratory environments.
02 / Hardware
Capture systems selected for the signal your model needs.
DatraAI deploys lightweight egocentric wearables, synchronized stereo systems and portable manipulation rigs. The final hardware configuration is matched to the task, operator, environment and required output.
Configurations shown are representative systems available through DatraAI data programs. Exact configuration and availability are confirmed during scoping.
01 / CAPTURE SYSTEMField-ready RGB + IMU
Monocular egocentric wearable
Best for scalable first-person human activity and manipulation capture.
A 30 g head-mounted camera module recording wide-field first-person RGB with tightly synchronized 9-DOF motion sensing. It supports live app-connected capture and standalone field recording.
Key specifications
- Recording: 1920 × 1080 at 30 FPS
- Horizontal field of view: 180°
- Motion sensing: 9-DOF IMU at 562.5 Hz
- IMU-frame synchronization delay: <1 ms
- Camera module weight: 30 g, excluding head strap
- Capture modes: live USB/UVC streaming or standalone recording
- Interface: USB-C / UVC
- Local storage: removable SD card, up to 2 TB
- Power: external power bank; up to 9 hours with the referenced 2,500 mAh configuration
Outputs
RGB video9-DOF IMUTimestampsCalibrationSession metadata
02 / CAPTURE SYSTEMGlobal-shutter dual RGB + IMU
Stereo egocentric wearable
Best for synchronized dual-view activity capture and geometry-aware datasets.
A lightweight binocular capture headset combining two synchronized global-shutter RGB cameras, high-rate IMU sensing, local storage, audio and wireless or wired data transfer.
Key specifications
- Camera system: 2 × RGB global-shutter cameras
- Recording: up to 1080p at 60 FPS
- Stereo baseline: 65 mm
- Field of view: up to 194° horizontal × 111° vertical per camera
- IMU sampling: up to 1 kHz
- Device weight: <150 g for the referenced configuration
- Battery runtime: >5 hours for the referenced built-in battery configuration
- Storage: removable MicroSD
- Connectivity: Wi-Fi 6, Bluetooth 5.4 and USB
- Audio: microphone capture supported
Outputs
Left RGBRight RGBHigh-rate IMUAudioTimestampsCamera intrinsicsCamera extrinsics
03 / CAPTURE SYSTEMPose + gripper + tactile + RGB
Portable visuotactile manipulation rig
Best for contact-rich, in-the-wild manipulation demonstrations.
A portable UMI-style capture system combining global-shutter visual sensing, visual-inertial pose estimation, gripper-state detection and high-resolution visuotactile force sensing.
Key specifications
- Vision: global-shutter RGB camera
- Vision field of view: 160°
- Frame rate: 30 FPS
- Localization: visual-inertial odometry using stereo fisheye cameras and IMU
- Tactile system: 2 × visuotactile sensors
- Tactile output: 3D force at 30 FPS
- Tactile resolution: up to 60,000 sensing points in the referenced configuration
- Stated tactile error: <5%
- Gripper-opening detection: magnetic encoder
- Position localization error: ≤3 mm in the referenced loop-closure test
- Attitude localization error: ≤0.3° in the referenced loop-closure test
- Time-synchronization error: ≤30 ms
- Continuous operation: ≥2 hours
- Backpack weight: 2.5 kg
- Output compatibility: LeRobot-compatible data format
Outputs
RGBDevice poseGripper opening3D tactile forceIMU / VIOTimestampsCalibration
03 / Actuate
Different sensors. One training-ready episode.
Actuate is DatraAI’s multimodal processing and quality layer. It verifies capture inputs, aligns sensor streams, recovers task structure, records provenance, applies quality and privacy gates, and packages usable episodes for the customer’s training stack.
Input / capture
Monocular RGBStereo RGBDepthIMUAudioPoseTactileRobot stateRobot action
Verify
Confirm the declared rig, channels, metadata and session integrity.
Synchronize
Align video, IMU, audio, pose, tactile and robot streams onto a shared timeline.
Understand
Recover hands, objects, depth, task intervals and action structure where supported.
Canonicalize
Map heterogeneous capture systems into a consistent, versioned episode representation.
Certify
Record data quality, processing status, provenance, consent and privacy results.
Export
Deliver LeRobot, RLDS, canonical or agreed customer schemas.
Output / delivery
Raw packageSynchronized sessionCanonical episodeLeRobotRLDSCustom schema
Engagement / Pilot to Scale
Start with a requirement, not a generic dataset.
A representative pilot establishes the capture configuration, quality bar and delivery structure before a program expands.
Define
Share the target tasks, environments, modalities, embodiment, volume, output format and acceptance criteria.
Pilot
DatraAI selects the capture configuration and runs a representative evaluation batch.
Validate
Your team reviews the signal quality, task coverage, schema and downstream usability before scale-up.
Scale
Collection expands through a documented operating plan, recurring quality checks and agreed delivery schedule.
04 / Research
Research that makes real-world data more useful for learning.
DatraAI studies how capture quality, multimodal alignment and physical representations affect learning from real interaction data. Research results, benchmarks and technical reports are published as they become ready.
Data quality and evaluation
Developing measurable links between capture quality, distribution coverage and downstream model performance.
Multimodal physical representations
Aligning RGB, stereo, depth, IMU, pose, tactile and robot signals into consistent representations of physical interaction.
World and action models
Exploring action-conditioned prediction and physical-world learning from real human and robot interaction data.