KAI World Model is a foundation world model developed by Kinetix AI, serving as an internal simulator/predictor for control policies on the KAI humanoid. The model learns the conditional distribution P(image_t+1, state_t+1 | image_t, state_t, action_t) from millions of hours of robot footage and synthetic data from simulators (Isaac Sim, MuJoCo). Once trained, it acts as a neural physics engine in which RL policies can be trained without using the real robot.
The architecture is based on a diffusion transformer (DiT) with conditioning state and cross-attention to language embeddings. The prediction resolution is 256×256 over 8 future timesteps. The model also generates an implicit contact-dynamics representation, crucial for object manipulation and locomotion on uneven terrain.
KAI World Model is a component of the proprietary Kinetix AI stack and is not publicly available. Internally it is used both for offline RL (on collected data) and for online imagination-based planning (Dreamer-style model-based RL). A research API release is announced for 2026.
A Runtime is the environment or execution layer used to run code, load libraries, manage dependencies, and operate applications or services — either in real time or during normal system operation. In robotics this includes real-time operating system (RTOS) runtimes, ROS 2 executor runtimes, containerised execution environments (Docker, podman), and embedded C++ runtimes on microcontrollers.
A Perception Stack encompasses the software layers that process data from cameras, LiDARs, IMUs, microphones, and other sensors in order to recognise the surrounding environment, perform localisation, detect and track objects, and interpret the scene. It is typically the first processing stage in an autonomous robot's data pipeline, feeding its outputs to planning and control stacks.
KAI humanoid (Kinetix AI) — internal use for training locomotion and manipulation policies. Publications at ICRA 2025 and CoRL 2025.
No open-source community. Three scientific publications in 2025 and a few conference demos.
Diffusion transformer ~3B parameters. One rollout generation ~120 ms on Thor.
License family: Proprietary – Commercial
Scaled to 3B parameters, added language conditioning.
First experimental version presented at ICRA 2025.
