RESEARCH

RECURSIVE CO-IMPROVEMENT

We advance world models, embodied agents, and AI infrastructure as a single co-evolving system. Our research closes the loop between virtual simulation and physical reality.

RESEARCH PILLARS

THREE FRONTIERS, ONE LOOP

01

World Models

  • ›Generative world models for physical environments
  • ›Editable, interactive, and persistent 3D scenes
  • ›Acceleration methods for real-time inference
  • ›Physical correctness and causal consistency
02

Embodied Agents

  • ›Vision-language-action models for robotics
  • ›Long-horizon planning and multi-agent collaboration
  • ›Skill abstraction and cross-embodiment transfer
  • ›Active exploration and failure-driven learning
03

AI Infrastructure

  • ›Large-scale reinforcement learning systems
  • ›Cross-world parallel rollout pipelines
  • ›Streaming inference and full-duplex execution
  • ›Heterogeneous hardware and edge-cloud coordination
RCI MECHANISM

CO-EVOLUTION, NOT ISOLATION

Traditional approaches train world models and agents independently. In RCI, each component reveals the weaknesses of the other: agents discover where the world model fails, and the world model generates harder, more realistic curricula in response.

1

Agent explores generated environment

Produces trajectories and identifies failure modes

2

Diagnostics pinpoint gaps

Attribution separates agent errors from environment flaws

3

Environment and agent update

Curriculum generation and policy improvement in parallel

4

External validation closes the loop

Real-world verification drives the next iteration

AGENT
WORLD
RCI
PUBLICATIONS & OPEN SOURCE

RESEARCH OUTPUT

World Model Acceleration & Generative Foundation Models

CVPR 2026 · TPAMI 2026

WORLD MODELS

Pre-training Infrastructure for Large Models

NSDI 2026 Best Paper Award

AI INFRA

Evaluation Frameworks for Embodied Intelligence

ICLR 2026

EVALUATION

Agent Training and Environment Construction

OSDI 2026

EMBODIED AGENTS

Large Model Inference Optimization

ATC 2025

INFERENCE
JOIN THE RESEARCH

SHAPE THE FUTURE OF PHYSICAL AI

We are building a research environment where world models, embodied agents, and systems evolve together. If you are excited by open problems at the frontier of physical intelligence, we want to hear from you.

EXPLORE PRODUCT