A data product system for embodied intelligence training. Human real-world operations, hand-object interactions, robot motion, and real-to-sim synthetic data are continuously processed into annotatable, trainable, reusable datasets and skill assets.
Anchored on first-person video as the master timeline, the eight modalities compose on demand — all linked through a unified episode, timestamps and coordinate frames.
01
TSING-E²CAP devices, third-person cameras and robot bodies capture human motion, hand-object interaction and environmental context.
02
Camera and sensor data is processed into parametric human and hand representations such as SMPL, SMPL-H, SMPL-X and MANO.
03
Task phases, action semantics, contact relations and object states are annotated to support fine-grained supervision for embodied models.
04
Real2Skill and sim2real-ready data are built with digital twins to extend coverage beyond real capture cost and distribution.
Following industry-leading data product logic: real-world task data is captured, standardized, and output as datasets and skill libraries compatible with LLMs and reinforcement learning.
High-precision SMPL / SMPL-H / SMPL-X human and hand motion library. Real human motion is parameterized into reusable body, hand and full-body pose representations for out-of-the-box motion libraries — powering humanoid priors, digital human driving, motion transfer and long-horizon task demos.
First- and third-person coverage of real scenes: observation paths, hand actions, object changes and environmental context. Built for vision models, task-phase recognition and embodied world models.
A high-precision MANO parametric database around fine hand actions, providing hand pose, finger motion and manipulation trajectory supervision for VLA, world model, EgoScale, VITRA and other training frameworks.
A fine-grained HOI dataset around hand-object contact: contact state, object 6D pose, manipulation phase and outcome state — supporting imitation learning and human policy modeling.
Multi-finger grasp, rotation, insertion, open/close, twisting, folding and other complex operations — focused on high-success-rate grasp policies and dexterous skill training for real objects and complex contact tasks.
Simulation-ready data for gait, balance, turning, obstacle avoidance, navigation and upper-limb coordination — for locomotion policy training and simulation validation.
Real task scenes are transformed into digital twin environments preserving objects, space, tasks and operation flows — for pre-deployment validation, policy stress testing and diversified sample expansion.
Robot-ready data packages for sim-to-real transfer that need no extra processing — helping teams shorten policy training, validation and deployment cycles.
Training-ready formats for VLA, world models, reinforcement learning and imitation learning: video, IMU, hand data, robot states, task metadata; human/hand parameters, object poses, contact relations, task phases.
Starting from raw multimodal capture, each layer matures into trainable datasets, reproducible simulation samples and transferable skill policies.
First/third-person multi-view raw footage that preserves the full scene and timeline — the source of every downstream asset.
Parametric human and hand representations (SMPL family, MANO) with semantic labels, ready for VLA and world-model training.
Digital-twin samples rebuilt 1:1 from real scenes — reproducible and extensible for Sim2Real transfer.
Task-level, transferable skill assets reused across embodiments through the unified WIOS interface.
Turning real human operations into a trainable, reusable, out-of-the-box data engine — reused across embodiments via the unified WIOS interface, ready to power model training.
