Multimodal Data Source
Supports synthetic, egocentric, UMI, and real-world teleoperation data. Compatible with heterogeneous multimodal inputs including vision, LiDAR, tactile sensing, motion states, and joint force/position signals.
00:00
Data Engine - LOOP
LOOP is a world-leading hybrid real–virtual data closed-loop system purpose-built for embodied AI and robotic learning. Deeply integrated with the AIRBOT product family, the DISCOVERSE simulation platform, and the ORION embodied foundation model, LOOP seamlessly supports diverse data modalities—including simulation-generated data, egocentric data, UMI, and real-world teleoperation. It establishes a fully unified, end-to-end pipeline spanning data acquisition, intelligent annotation, large-scale synthesis, model training, validation, and deployment. By breaking down the boundary between real and synthetic data, LOOP acts as a high-efficiency data engine that continuously drives model iteration. It not only powers scalable data production but also accelerates model development—systematically constructing the embodied data pyramid while significantly improving generalization and robustness. Every data cycle becomes a catalyst for intelligence evolution.
Unlike traditional AI, embodied AI learns through interaction with the physical world. However, such interaction data is difficult to collect at scale, and high‑quality embodied data has become the core bottleneck limiting the advancement of foundation models. LOOP deeply understands the relationship between multimodal data across the embodied data pyramid and task-oriented model capabilities. It provides a complete data loop for all system layers of the ORION embodied foundation model, serving as an efficient engine for data acquisition and robotic intelligence evolution. LOOP enables S3 to leverage large-scale internet data, enhancing broad cognition and generalization priors through data mining, automated cleaning, and foundation model distillation. With a complete, efficient, low-cost hybrid data acquisition system, LOOP precisely routes data flows—providing S2 and S1 with egocentric environmental understanding and end-effector interaction data, bridging the gap between simulation and reality. Power by high-performance standards and EOL Calibration, LOOP ensures data consistency and quality. High-frequency motion control data is directly delivered to S0, enabling real-time response, human-like agility, and uncompromised safety in real-world deployment. Through this architecture, LOOP empowers ORION with top‑layer cognitive breadth, middle‑layer perceptual depth, and bottom‑layer reflex‑level speed. With continuous forward data flow and backward validation across layers, LOOP drives ORION toward self-evolving intelligence in the physical world.
Key Features
Real2Sim2Real Pipeline
Covers the Real2Sim2Real pipeline —— From data acquisition and annotation to synthesis, training, evaluation, and deployment. A unified platform eliminates tool fragmentation and enables seamless data–model iteration.
AI-Empowered Data Engine
AI is deeply embedded across intelligent annotation, quality control, hard-case mining, and simulation generation. Through active learning and automated optimization, LOOP continuously improves data efficiency and accelerates model performance gains.
Compatible Hardware
AIRBOT SC0Ego-Centric Device
AIRBOT SC2Handheld Data Acquisition Device
AIRBOT PlayBimanual Data Acquisition Device
