kairos-agi/kairos

What it solves

Kairos is a 4B-parameter native cross-embodiment world model designed to bridge the gap between passive video understanding and active robotic control. It addresses challenges such as heterogeneous embodiment data, weak long-horizon reasoning, and the high compute requirements of edge-side deployment in embodied AI.

How it works

Kairos uses a unified Mixture-of-Transformers framework to jointly handle understanding, generation, and action prediction. It is trained using a Cross-Embodiment Data Curriculum (CEDC) that progressively moves from general videos to human behavior and real-robot interaction data. To enable efficient long-horizon modeling and real-time deployment, it employs a hybrid linear temporal memory mechanism combining Sliding-Window, Dilated Sliding-Window, and Gated Linear Attention.

Who it’s for

It is intended for researchers and developers working on embodied intelligence, robotics, and world models who need a model that can generalize across different hardware platforms (such as Agibot G1, Unitree G1, and Songling PIPER) and predict executable action trajectories.

Highlights

• Cross-Embodiment Generalization: A single model that works across single-arm, dual-arm, and dexterous-hand platforms.
• World-Action Model: Predicts executable robot action trajectories directly from visual observations and task context.
• Efficient Edge Deployment: Optimized for low-latency, high-throughput inference on edge-side GPUs (e.g., RTX 5090, A800).
• Physical-Causal Consistency: Uses causal Chain-of-Thought (CoT) to transform multimodal inputs into deep task logic for autonomous planning.