Wearable multimodal collection
- First-person video
- Spatial localization
- Eye movement and voice
- Hand motion and context
SceneAct AI
Robot-ready task data from egocentric reconstruction
SceneAct AI reconstructs real human workflows from first-person multimodal capture into episode data for robot imitation, VLA training, task understanding, and evaluation.
Wearable AR glasses capture first-person video, spatial localization, eye movement, hand motion, voice, and context from natural daily tasks.
View OutputRobot-ready output
Episode data assets
Reconstructed spatial interaction aligned with SOP steps, action primitives, outcomes, anomaly tags, and quality scores.
SceneSpatial structure
HandTrajectory and grasp
TaskPhase and outcome
Egocentric interaction reconstruction
Multimodal models recover the task state that raw first-person video does not provide: structure, hand-eye trajectories, object states, grasp events, action phases, and optional full-body posture.
- Reconstruct scene structure from egocentric observations
- Recover hand-eye trajectories and object states
- Align interactions with SOP steps and action primitives
Product modules
From real tasks to trainable robot data
SceneAct combines wearable capture, egocentric reconstruction, and robot-ready episode generation into one data engine.
Spatial interaction engine
- Scene structure
- Hand-eye trajectories
- Object states
- Grasp events and phases
Robot-ready episodes
- SOP alignment
- Action primitives
- Success and anomaly labels
- Quality-scored assets
Use cases
Build task data before robots enter the room
Early focus areas include medical eldercare, preclinical labs, smart laboratories, and high-standard industrial operations.
Why these workflows
Clear procedures, frequent tasks, stable tools, and scarce robot training data.
[email protected]
EMAIL
[email protected]