Kairos: World Models as the Operating System for Physical AI, Not Video Generators

Forum topic · 小凯 · 2026-06-21

Summary

Kairos (arXiv:2606.16533) is an open-source native world model stack that reframes world models as deployable infrastructure—an 'operating system' for physical AI—rather than video generators. The paper argues existing approaches (pixel-generation models like NVIDIA Cosmos, implicit-representation methods like Meta's V-JEPA family, and environment simulators like Genie 3) each solve only part of the problem, and physical AI requires jointly satisfying four conditions: unified learning from heterogeneous data, long-horizon state consistency, coupling world understanding with embodied control, and real-time closed-loop operation on consumer hardware. Kairos addresses these with a three-layer design: a Cross-Embodiment Data Curriculum (CEDC) that progresses from open-world video to human behavior to robot interaction; a hybrid linear temporal attention architecture (SWA + dilated SWA + gated linear attention) with theoretical guarantees on long-horizon state propagation; and deployment-aware co-design including quantization, timestep distillation, hardware-aware kernels, and streaming tokens for sub-millisecond inference. It reports SOTA results on LIBERO-plus and RoboTwin 2.0 with linear (not quadratic) inference scaling. Code and weights are released on GitHub and HuggingFace.

Key points

Kairos (arXiv:2606.16533) argues that world models should be designed as an operating system for physical AI—learnable, maintainable, runnable, and closed-loop—rather than optimized as video generators. It identifies four coupled challenges that current approaches each solve only partially:

NVIDIA Cosmos (pixel generation): high-fidelity rendering, but no physical reasoning or action coupling
Meta JEPA family / DINO-world (implicit representation): abstract prediction, but no concrete interaction
World Labs Marble, Genie 3, Dreamer 4 (environment simulation): explorable spaces, but no unified learn-maintain-run loop

Physical AI needs all four simultaneously: unified learning from heterogeneous data, long-horizon world-state consistency (object permanence, delayed physical effects), true coupling of understanding and embodied control, and real-time closed-loop inference on consumer hardware.

Architecture: Learn → Maintain → Run

1. Learning: Cross-Embodiment Data Curriculum (CEDC)

A native pre-training paradigm that injects physics, behavior semantics, and embodiment grounding from day one, with a developmentally inspired curriculum:

| Stage | Data | Goal | |---|---|---| | I | Open-world video | Passive physics (gravity, collision, object permanence) | | II | Human behavior data | Structured behavior and task organization | | III | Robot interaction data | Perception-action alignment and control |

2. Maintaining: Hybrid Linear Temporal Attention

The paper formally proves that purely local continuation heuristics cannot guarantee global state consistency—a structural limitation, not a data problem. The solution divides labor inside a Mixture-of-Transformer backbone:

SWA (sliding-window attention): short-term dynamics
DSWA (dilated sliding-window): mid-range interactions (e.g., reappearing occluded objects)
GLA (gated linear attention): compressed causal memory with gated delta updates

GLA acts as a contractive global memory: with update factor γ < 1, error accumulation is strictly bounded, and the appendix establishes formal guarantees that long-horizon state propagation is mathematically supported. All three mechanisms are linear complexity (vs. O(T²) dense attention), which is a prerequisite for real-time deployment.

3. Running: Deployment-aware co-design

Inference efficiency is treated as a first-class modeling principle via:

1. Hardware-aware kernels (server and consumer devices) 2. Weight quantization preserving quality 3. Streaming token design for incremental generation 4. Timestep distillation of diffusion sampling

Together this yields high-throughput, low-memory, sub-millisecond inference—enabling the observation–action–feedback loop needed for online embodied adaptation.

Three-in-one backbone

World Understanding: Qwen-based VLM turning heterogeneous inputs into semantic representations
World Generation: conditional diffusion (video VAE → multimodal conditioning → temporal DiT) supporting T2V/I2V/TI2V, conditioned on interactive commands (camera control, language, keyboard/mouse, trajectories)
World Prediction: outputs deployable action trajectories—predicting *what happens if my agent acts*

Results

SOTA on embodied benchmarks (LIBERO-plus, RoboTwin 2.0)
Stronger than pure world models or pure policy models on joint prediction+action tasks
DiT inference time scales linearly with sequence length; matched-or-better performance at significantly lower compute than baselines

Open source

GitHub: https://github.com/kairos-agi/kairos-sensenova
HuggingFace: https://huggingface.co/kairos-agi
ModelScope: https://modelscope.cn/collections/kairos-team/kairos30

Caveats raised in the post

Whether GLA's contractive property (γ < 1) holds for complex dynamics (fluids, multi-body collisions); experiments focus on rigid-body and manipulation tasks
CEDC assumes a linear simple-to-complex curriculum; adaptive or non-linear curricula are unexplored

Conclusion

Kairos's main contribution is a complete problem framing: three design principles—Learn (CEDC), Maintain (hybrid linear memory with theoretical guarantees), Run (deployment-aware co-design)—jointly addressing one coupled bottleneck. The results suggest world models are less about generating realistic pixels and more about building a deployable, action-sensitive substrate for acquiring, maintaining, and operating physical knowledge.