Manifold AI Research

We build world models and robot policies that connect perception, reasoning, and action.

Our research focuses on embodied intelligence: enabling robots to understand physical environments, learn from experience, and perform long-horizon tasks under multimodal guidance.

Research Focus

• Interactive World Models: Simulating physical environments through controllable, interactive generation.
• World Action Models: Jointly modeling future observations and robot actions for multimodal control and in-context learning.
• Recursive Robot Self-Improvement (R2SI): Improving robot skills and policies through experience, reflection, and iteration.

Featured Projects

WorldScape-MoE

A unified mixture-of-experts world model for scalable heterogeneous action control, sharing world knowledge across navigation, robotic manipulation, and egocentric hand interaction.

Code · Project Page

WorldScape Policy 2.0

A world-action policy combining reasoning-augmented memory and event-grounded pretraining for multimodal robot control.

Models · Code · Project Page

WorldScape R²SI

A robot agent with an autonomous self-improvement loop that refines reusable skills, memory, and policies through collected experience and validation.

Models · Code · Project Page

Connect

Website · GitHub