ACE Robotics Unveils Full-Stack Physical AI Breakthrough with Kairos 3.1 at WAIC

The WAIC 2026 World Model "Six Little Dragons" summit in Shanghai gathered six leading embodied‑intelligence companies. ACE Robotics, dubbed the "Roll King" for its rapid development, announced three core breakthroughs: the Kairos 3.1 world model, the Environment‑centric Data Collection 2.0 pipeline, and a suite of vertical solutions.

Kairos 3.1 extends the earlier Kairos 3.0 (Kairos) by fusing generative, physical, and cognitive intelligence into a unified "understand‑generate‑predict" loop. The model uses a hybrid Transformer architecture with shared attention to compress multimodal inputs into a dense latent space, enabling seamless transition from perception to action planning.

In the perception stage, ACE‑BRAIN‑0.5 provides spatial understanding, having secured twelve global SOTA benchmark rankings. For generation, the model builds a high‑fidelity parallel digital world containing ~300 k Chinese residential floor plans, 5 000 full‑house simulation scenes, and 8 700 high‑quality 3D assets, each annotated with geometry, material, and physics properties. The prediction stage lets robots simulate multiple candidate trajectories before execution, dramatically reducing trial‑and‑error costs.

Edge deployment is a key focus: on the NVIDIA Jetson Thor platform with BF16 precision, Kairos 3.1 8B achieves an average inference latency of 125 ms, matching classic VLA model speeds while supporting on‑device reasoning without cloud round‑trips.

ACE Robotics also introduced the Information Density Law for embodied intelligence, categorising data from L1 to L5 based on richness—from monocular video (L1/L2) to high‑resolution force‑touch and continuous interaction data (L5). This law guides the design of the Environment‑centric Data Collection 2.0 suite, comprising the ACE Ego Kit (head, hand, chest sensors with 0.01 N force sensitivity), ACE Data Engine (automatic labelling and a 4D hand‑capture method achieving 0.997 frame‑level accuracy), and ACE Ego Matrix (cross‑embodiment data unification).

Three vertical solutions were demonstrated: Xiaoman (instant‑retail fulfillment robot W1, 75 cm minimum aisle width, deployed in Sense MartGo and KuaiKeDa), Xiaoxin (hotel laundry automation with full‑process handling and failure‑aware replanning), and Xiaotu (open‑scene autonomous robot dog operating 24/7 for security and facility inspection). All solutions rely on the Kairos 3.1 loop—understand, plan, act, evaluate, and self‑reflect—to achieve continuous improvement in real‑world tasks.

The article concludes that world models are moving beyond pure video generation toward integrated control loops, and that ACE Robotics’ speed, end‑to‑end data pipeline, and industry deployments position it as a full‑stack player in the physical AI race.