Kunlun Spins Out Riemann Dynamics, Unveils World Action Model for Robotics

0xAllen_ · x · 2026-07-21

Riemann Dynamics, an embodied AI team incubated within Kunlun, has spun off and released its World Action Model, Riemann-1.0. Originally focused on game world models, the team pivoted to robotics after discovering that real-world data significantly improved the model's physical understanding.

Core Tech & Data

World Action Model: Unlike traditional VLA models that directly output actions, this model predicts environmental changes caused by actions and continuously updates its state. This enables long-horizon tasks, cross-scene generalization, and cross-embodiment transfer.
Data Strategy: Moving away from repetitive factory data, the team focuses on human egocentric (first-person) videos. They have amassed 232,000 hours of multi-source data (including 200k hours of first-person video) across 41 robot embodiments, aiming to scale to 1 million hours.

Commercialization & Vision

Target Scenarios: Avoiding highly standardized factory pipelines, the team targets environments rich in long-tail problems like homes, hotels, and nursing facilities, with initial paid demand seen in Japan.
Ultimate Goal: Building a universal robotic API allowing hardware vendors to plug in without retraining models. The founder predicts a "GPT-3 moment" for embodied AI by late 2027 or early 2028, shifting the industry from selling hardware to selling services.

Original post →

Kunlun Spins Out Riemann Dynamics, Unveils World Action Model for Robotics

More from Companies & People