RESEARCH AT SKYWORK AI

Fei Kang 康飞

World models.
Embodied intelligence.

My work spans model architecture, training, post-training, and training data development for world models and embodied foundation models, including World Action Models (WAMs).

PreviouslyTencent3D Engine Technical Expert

WORLD / STATE / ACTION
Matrix-Game 3.0Real-time interactive world generation
↗
Official project demonstration00:15

01 / SELECTED WORK

Selected research.

From interactive virtual worlds to executable robot policies.

6 papers

More research

2026

Matrix-Game 3.0

Co-author. Data development and model architecture contributions. Memory-augmented world generation at up to 720p / 40 FPS with the 5B model.

2025

Matrix-Game 2.0

Co-author. Designed and built the Unreal Engine data-generation pipeline. Real-time action-conditioned generation at 25 FPS.

2025

Matrix-Game

Co-author. A 17B interactive world foundation model for controllable Minecraft generation.

02 / FEATURED CONTRIBUTION

A data engine
for world models.

I pioneered the use of Unreal Engine to generate precisely synchronized action–video pairs for interactive world-model training, designing and building the pipeline introduced in Matrix-Game 2.0.

Read the technical record ↗

Introduced in Matrix-Game 2.0 · 2025

03 / BACKGROUND

From 3D engines
to world models.

Connecting experience in 3D engines and simulation with model research and training.

CURRENT

Skywork AI

World models & embodied AI

Model architecture, training, post-training, and training data development for world models, embodied foundation models, and WAMs.

PREVIOUSLY

Tencent

3D Engine Technical Expert

A background in 3D engine technology, real-time systems, and simulation that informs my research on interactive worlds.