ORION is a world-leading hierarchical embodied AI foundation model, pioneering a four-tier architecture of S3–S2–S1–S0.
ORION’s layered design precisely separates intelligent decision-making from physical safety along both temporal sensitivity and functional hierarchy. S3 (Slow Brain) focuses on non-real-time deep reasoning and long-term memory. It enables complex semantic understanding and continuous knowledge accumulation. S2 (Fast Brain) handles real-time spatial perception, scene understanding, and action prediction, allowing rapid planning in dynamic environments. S1 (Cerebellum) is responsible for real-time perception and control execution, generating action commands that directly interact with the physical world. S0 (Peripheral Layer) serves as the safety-critical foundation, enforcing hardware-level constraints such as collision avoidance, torque limits, and emergency braking—ensuring safe execution regardless of higher-level decisions.
ORION adopts a highly flexible dual-track training paradigm to accommodate diverse application scenarios and cost constraints. Decoupled Optimization: S0 and S3 can be independently trained and iterated. S0 leverages classical robotics principles combined with lightweight perception models. S3 benefits from large-scale multimodal foundation models pretrained on internet-scale data. Embodied Core (S1+S2) Training Paths: Hierarchical Lightweight Training requires only small-scale datasets to quickly fine-tune deployable skills, significantly reducing data acquisition costs and deployment time. End-to-End Full-Scale Training uses large-scale, multi-source data for joint optimization, achieving stronger generalization in open and complex environments. Crucially, as S3 continues to evolve, it can distill knowledge into S1+S2, continuously expanding the action space—from limited real-world data to an effectively unbounded imagination-driven reasoning space—enabling the self-evolution of physical intelligence.
Integrating vision, language, and tactile signals, it builds an interpretable and highly generalizable world model. Beyond black-box end-to-end mapping, it explicitly understands physical dynamics to enable reliable decision-making in open environments.
Enabling accurate spatial perception through explicit 3D reconstruction rather than monocular estimation. This significantly improves pose estimation across diverse object shapes and sizes, enhancing manipulation robustness and reducing failure rates.
Enabling online reinforcement learning during inference, continuously optimizing policies through execution feedback. Robots evolve through real-world experience without retraining.
Designed for seamless integration with the S0 layer, it incorporates advanced control protocols and spatiotemporal memory to smooth trajectories and compensate for latency. This ensures stable, precise, and safe real-time execution despite delays in high-level model inference.
