Lead of Frontier Technology Exploration · Spirit AI
I lead Frontier Technology Exploration at Spirit AI. I received my Ph.D. from the Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University, advised by Prof. Yang Gao. Previously, I obtained my bachelor's degree from Beijing University of Posts and Telecommunications (BUPT).
My research focuses on Embodied AI, a frontier domain at the intersection of machine learning, robotics and computer vision. My goal is to develop general-purpose robots endowed with human-like agility and dexterity, and capable of generalizing learned behaviors to novel tasks and unseen real-world environments.
My work builds toward what I call the Keystone Arch Architecture: two pillars, locked together by a keystone.
A large-scale robot policy that unifies the modules of the stack in a single architecture and feeds on the data engine — generalizing broadly across tasks and embodiments.
Renovating the classical perception → planning → control pipeline by embedding scalable learning-based methods at every layer.
There is no “Internet for Robots.” I study which data actually matters for physical intelligence, and how to scale it.
We are organizing the Data-Centric Robotics Workshop at RSS 2026.
We release OpenHLM, an empirical recipe for whole-body humanoid loco-manipulation.
VLAs need to reason — but they need to know when to reason (or not)! We release OneTwoVLA.
We release Data Scaling Laws in Imitation Learning for Robotic Manipulation.
Semantic-Geometric Representation (SGR) is accepted at CoRL 2023.
Our work on pre-trained vision models in motor control is accepted at ICML 2023.
SFC is accepted at ECCV 2022 as an oral presentation!
Representative papers (* denotes equal contribution). For the full list, see my Google Scholar.
Yingdong Hu*, Haodong Zhu*, Boyuan Zheng*, Yihang Hu*, Tong Zhang*, Zunhao Chen, Junming Zhao, Ruiqian Nai, Yang Gao
arXiv preprint, 2026
A recipe for whole-body humanoid loco-manipulation, turning a language instruction and the robot's observations into coordinated motion across every degree of freedom.
Fanqi Lin*, Ruiqian Nai*, Yingdong Hu*, Jiacheng You, Junming Zhao, Yang Gao
International Conference on Learning Representations (ICLR), 2026
We introduce OneTwoVLA, a single unified vision-language-action model capable of both acting (System One) and reasoning (System Two). Crucially, it adaptively determines when to engage each mode.
Yingdong Hu*, Fanqi Lin*, Pingyue Sheng, Chuan Wen, Jiacheng You, Yang Gao
International Conference on Learning Representations (ICLR), 2025
We demonstrate that a policy's generalization ability to new objects, new environments, or both scales approximately as a power law with the number of training objects, training environments, or training environment–object pairs, respectively.
ByteDance Seed
Technical Report, 2025
GR-3 is a large-scale vision-language-action (VLA) model. It showcases exceptional capabilities in generalizing to novel objects, environments, and instructions involving abstract concepts.
Yingdong Hu*, Fanqi Lin*, Tong Zhang, Li Yi, Yang Gao
Workshop on Vision-Language Models for Navigation and Manipulation, ICRA, 2024
We introduce ViLa, a novel approach for long-horizon robotic planning that leverages GPT-4V to generate a sequence of actionable steps, empowering robots to execute complex tasks with a profound understanding of the visual world.
Tong Zhang*, Yingdong Hu*, Hanchen Cui, Hang Zhao, Yang Gao
Conference on Robot Learning (CoRL), 2023
We present Semantic-Geometric Representation (SGR), a universal perception module for robotics that leverages the rich semantic information of large-scale pre-trained 2D models and inherits the merits of 3D spatial reasoning.
Yingdong Hu, Renhao Wang, Li Erran Li, Yang Gao
International Conference on Machine Learning (ICML), 2023
We conduct the first thorough evaluation of pre-trained vision model performance across different downstream policy learning methods and environments, discovering that the effectiveness of pre-training is highly dependent on the choice of the downstream policy learning algorithm.
Yingdong Hu, Renhao Wang, Kaifeng Zhang, Yang Gao
European Conference on Computer Vision (ECCV), 2022
We show that fine-grained features learned with pixel-level self-supervised learning (SSL) objectives are complementary to semantic features from image-level SSL methods. Fusing these features can significantly improve performance on visual correspondence tasks.
