[2606.16533] Kairos: A Native World Mannequin Stack for Bodily AI

View a PDF of the paper titled Kairos: A Native World Mannequin Stack for Bodily AI, by Kairos Workforce: Fei Wang and 22 different authors

View PDF
HTML (experimental)

Summary:World fashions are transitioning from passive visible turbines to foundational, operational infrastructure for Bodily AI: they have to natively purchase world data from heterogeneous expertise, preserve persistent states over lengthy horizons, and execute effectively inside actual deployment constraints. We introduce Kairos, a local world mannequin stack designed round these necessities. (1) Kairos learns the world by pioneering a Native Pre-training Paradigm ruled by a Cross-Embodiment Knowledge Curriculum, which organizes open-world movies, human behavioral information, and robotic interactions right into a progressive developmental pathway. (2) Kairos maintains the world by unified world understanding, era, and prediction inside a Native Unified Structure outfitted with Hybrid Linear Temporal Consideration, the place sliding-window consideration captures native dynamics, dilated sliding home windows seize mid-range dependencies, and gated linear consideration maintains persistent international reminiscence. We set up formal theoretical bounds demonstrating that this temporal factorization strictly limits error accumulation, mathematically guaranteeing state propagation throughout prolonged horizons. (3) Kairos runs the world by incorporating a Deployment-Conscious System Co-Design to help low-latency rollout era on server and consumer-grade {hardware} for real-world observation-action-feedback loops. Experiments on embodied world-model, long-horizon, and action-policy benchmarks present that Kairos achieves high degree efficiency whereas providing a powerful efficiency-capability trade-off. Collectively, these outcomes place Kairos as a cohesive operational basis for future self-evolving bodily intelligence.