Industry Experience

I work across data, models, and real-robot deployment, with an emphasis on turning embodied-AI research into systems that operate reliably in the physical world.

VLA pretraining and data pipelines

Processed large-scale real-robot datasets, including AgiBot-World and DROID, through action normalization, tokenization, and structured conversion for cross-embodiment VLA pretraining.

RoboBrain-X0

VLA reinforcement post-training

Reproduced and evaluated robot-policy post-training pipelines on real-robot manipulation tasks, contributing experiments to FORCE.

FORCE

RL post-training · USB insertion

RL post-training · Cup pickup

Biological-laboratory VLA deployment

Trained and deployed VLA policies, including LingBot-VLA, π₀, and π_0.5, for laboratory automation. Integrated perception, policy inference, and control on real robot platforms, followed by iterative evaluation and failure analysis.

Laboratory manipulation · Demo I

Laboratory manipulation · Demo II

World-model representation learning

Developing representation-learning methods and data-cleaning pipelines for foundational world-action-model pretraining. I also organize a collaborative knowledge base on representations for world models.

Awesome Representation for World Models

Physical world model · Tabletop curling

Built a closed-loop demo in which a robot observes an evolving game state, predicts physical outcomes, and competes against a human player.

Physical world model · Human–robot tabletop curling

Industry Experience

Beijing Academy of Artificial Intelligence

VLA pretraining and data pipelines

VLA reinforcement post-training

XtalPi

Biological-laboratory VLA deployment

Muka Robotics

World-model representation learning

Physical world model · Tabletop curling