How Daxiao’s Kairos Beats Nvidia and Redefines Physical AI with a Native Integrated World Model

In December 2025 Daxiao Robot released the Kairos "multimodal understanding — generation — prediction" native integrated architecture; by March 2026 the system was fully deployed on edge devices. Nvidia Cosmos 3.0 later adopted the same underlying design.

Kairos achieved top rankings on four authoritative embodied‑intelligence benchmarks—RoboTwin 2.0, LIBERO‑Plus, WorldModelBench, and DreamGen Bench—demonstrating a paradigm shift in world‑model technology.

Three mainstream world‑model streams

Pixel‑level generative rendering (e.g., Nvidia Cosmos) synthesizes high‑fidelity video but incurs heavy compute cost and remains a content‑creation tool.

Interactive environment modeling (e.g., DeepMind Genie 3, Dreamer series) builds persistent simulators for long‑horizon planning but relies on recursive imagination.

Predictive latent‑representation learning (e.g., Meta JEPA) learns abstract physical structures for zero‑shot planning and control, offering better computational efficiency.

All three paths share bottlenecks: misalignment between visual semantics and robot actions, error accumulation across staged pipelines, and limited role beyond content generation.

Native integrated architecture

Kairos introduces a single Mixed‑type Transformer (MoT) that fuses world understanding, generation, and prediction into one shared backbone. The MoT maintains a global state via a mixed linear temporal‑memory mechanism, eliminating the representation gap and cumulative drift of modular designs.

The understanding module extracts physical laws, causal logic, and task semantics from heterogeneous data, providing precise semantic anchors. The generation module transforms these anchors into physically consistent environment dynamics, serving strategy simulation rather than pure visual output. The prediction module, co‑designed with generation, directly outputs executable robot action trajectories; it can also run in a pure‑action mode that skips video synthesis, improving accuracy and latency.

Cross‑embodiment progressive training (CEDC)

Physical pre‑training : leverages millions of hours of open‑world video covering gravity, collisions, fluid dynamics, etc., across four core domains (human, robot, generic scenes, physical phenomena) to build foundational world knowledge.

ACE human‑centered data : incorporates ~100 k hours of high‑precision human‑operation recordings, preserving task intent, tool use, and household activities, thereby bridging the semantic gap between perception and control.

State‑action joint training : integrates high‑quality robot interaction datasets (dual‑arm collaboration, flexible manipulation) to tightly align perception representations with action spaces.

This hierarchical data pyramid enables cross‑embodiment learning, preserving the broad generalization of massive data while achieving precise robot‑control anchoring.

Edge‑side native deployment

Algorithmic distillation compresses the diffusion process from dozens of steps to four using a “distribution matching + consistency constraint” framework, retaining physical fidelity. Hardware co‑design applies mixed‑parallel scheduling, DiT feature caching, operator fusion, FP8 low‑precision arithmetic, INT4 weight quantization, and block‑wise streaming memory to reduce memory footprint and latency.

Empirical results show that a 480p, 5‑second physics‑simulation video is generated in 3 seconds on a 4 × A800 GPU cluster, delivering a 2.5–3.7× speedup over comparable models and up to 85× faster than billion‑parameter universal world models.

Technical report links: https://arxiv.org/abs/2606.16533 and https://huggingface.co/papers/2606.16533