DAXIAORobotics/kairos

What it solves

Kairos addresses the limitations of current embodied AI world models, specifically the gap between video generation and executable robot actions. It solves the problem of heterogeneous embodiment data and weak long-horizon reasoning, allowing a single model to understand physical laws, generate plausible future visual states, and predict precise robot action trajectories across different hardware platforms.

How it works

Kairos is a 4B-parameter native cross-embodiment world model based on a Mixture-of-Transformers framework. It utilizes a Cross-Embodiment Data Curriculum (CEDC) to learn from general videos, human behavior, and real-robot interactions. To handle long-horizon modeling efficiently, it employs a hybrid linear temporal memory (combining Sliding-Window, Dilated Sliding-Window, and Gated Linear Attention) to maintain persistent world states. This allows the model to jointly predict future visual frames and executable actions.

Who it’s for

This project is for researchers and developers in embodied AI and robotics who need a foundation model capable of zero-shot multi-task generalization across various robot platforms (such as Agibot G1, Unitree G1, and Songling PIPER).

Highlights

• Unified Architecture: Integrates understanding, generation, and action prediction into a single intelligence loop.
• Cross-Embodiment Generalization: A single "brain" that works across single-arm, dual-arm, and dexterous-hand platforms.
• Efficient Deployment: Optimized for real-time edge-side inference with low memory and high throughput.
• Physical-Causal Consistency: Uses causal Chain-of-Thought (CoT) to move from simple command execution to intent understanding.