Zhuoyang Liu | 刘卓洋

Hi there! I'm a third-year undergraduate student at Peking University, majoring in Computer Science. Currently, I am a visiting student at BAIR at UC Berkeley, advised by Prof. Trevor Darrell. Prior to this, I was a research intern at the HMI Lab at Peking University, advised by Prof. Shanghang Zhang. My research interests lie in the application of multimodal large models, specifically VLA models, in robot manipulation.

Email  /  Scholar  /  Twitter  /  Github  /  WeChat

Research

I'm interested in Computer Vision, Robot Learning and Embodied Large Multimodal Models. My research focuses on how to enhance large multimodal models to better reason about the physical world and develop effective task planning. Some papers are highlighted.

TwinRL is a digital twin-real-world collaborative RL framework designed to scale and guide exploration for VLA models.

A VLA model that enables efficient reasoning before acting through a Latent Spatio-Temporal Chain-of-Thought (CoT), capturing fine-grained physical and robotic dynamics that are often difficult to verbalize.

A unified VLA framework built upon a Mixture-of-Transformers (MoT) architecture, enabling coherent collaboration between multimodal manual generation and action execution.

DualVLA improves action performance through carefully designed post-training while preserving the reasoning ability.

A multisensory language-action (MLA) model that collaboratively perceives heterogeneous sensory modalities and predicts future multisensory objectives to facilitate physical world modeling.

RoboMIND 2.0 is a comprehensive real-world dataset comprising over 310K dual-arm manipulation trajectories collected across six distinct robot embodiments and 739 complex tasks.

Adaptive Coordination Diffusion Transformer (AC-DiT) enhances mobile base and manipulator coordination for end-to-end mobile manipulation.

Unlike previous dual-system VLA methods that attach a separate policy head as System 1, FiS-VLA repurposes the final transformer blocks of an intact VLM as System 1, while retaining the full model for System 2 reasoning.

3DS-VLA enhances pretrained 2D vision-language models (VLMs) with comprehensive 3D awareness, enabling the prediction of robust end-effector poses.

HybridVLA innovatively integrates diffusion and autoregressive action prediction within a single LLM, fully leveraging the continuity and probabilistic nature of diffusion alongside the reasoning capabilities of autoregressive modeling.

H2R is a simple data augmentation technique for robot pre-training from videos, which extracts the human hands from first-person videos and replaces them with those of different robots to generate new video data for pre-training.

Research on Embodied AI and Robot Manipulation

Focused on Vision-Language-Action (VLA) models

Embodied AI
Research Advisor: Prof. Shanghang Zhang

This homepage is designed based on Jon Barron's website and deployed on Github Pages.