白辰甲
Research Scientist
Institute of Artificial Intelligence (TeleAI), China Telecom
Biography
I am a Research Scientist at Institute of Artificial Intelligence (TeleAI), China Telecom and the Director of Embodied AI research center, specialized in the cutting-edge field of Embodied AI and Reinforcement Learning (RL). Our group is dedicated to develop embodied technologies encompassing perception, planning, locomotion, manipulation, and promoting the industrial application of embodied AI. Our group thrives under the leadership of Prof. Xuelong Li, who serves as the dean of TeleAI. Previously, I was a Researcher at Shanghai AI Laboratory, affiliated with IPEC group. My research interests include diffusion/transformer policy, LLM-driven planning, world model, preference learning, RL/MPC-based locomotion, dexterous manipulation, representation learning, sim-to-real, multi-agent collaboration, as well as real-world applications for robot arm, dexterous hand, quadruped robot, and humanoid robot.
I holds a Ph.D. degree in Computer Science from Harbin Institute of Technology (HIT), advised by Prof. Peng Liu. I am fortunate to have been collaborated with many fantastic researchers. I was a joint PhD student at University of Toronto and Vector Institute, working with Prof. Animesh Garg. I also used to be an intern at Huawei Noah’s Ark Lab (advised by Prof. Jianye Hao), Tencent Robotics X (advised by Dr. Lei Han), and Alibaba. I received my Bachelor’s degree and Master’s degree in Computer Science from HIT.
中文简介：白辰甲，博士，现任中国电信人工智能研究院（TeleAI）研究科学家、具身智能研究中心主任，兼任清华大学、上海交通大学、复旦大学行业导师，联合培养专项工程博士。入选第十届中国科协青年“托举”人才，上海市青年科技英才“扬帆计划”，上海市徐汇区“光启”青年人才。负责研发TeleBot系列人形机器人和轮式机器人研发，在技术上构建软硬件一体化、大小脑协同等具身智能体系，实现具身智能体的通用策略学习，推动机器人应用。围绕具身大脑构建了跨本体适配模型、灵巧操作任务规划平台、和数据合成平台；围绕具身小脑构建了首个开源机器人全身仿人运动控制框架、文本驱动的通用小脑、感控一体人形机器人等。学术成果方面，已发表高水平论文80余篇，包括机器学习顶会NeurIPS、ICML、ICLR，人工智能顶刊AI Journal、TPAMI、SCIC，机器人顶会ICRA、CoRL等，撰写《强化学习：前沿算法与应用》专著，由机械工出版社出版。撰写了国内首篇《大模型驱动的具身智能：发展与挑战》综述，下载量2万余次，获中国科学年度最具人气论文。承担国家自然科学基金、国家重点研发计划课题、上海市科委项目、中国电信内部立项等项目。获世界人工智能大会优秀论文提名奖、ICCV多地形人形机器人挑战赛冠军、哈尔滨工业大学优秀博士论文奖等奖项，相关成果受到MIT Technology Review、CCTV等媒体报道。担任顶级会议NeurIPS、ICML、AAMAS、ICME、PRCV领域主席，并担任多个顶级期刊和会议的审稿人。
团队招收具身智能方向全职研究人员、实习生、联培博士生，具体详见链接.
Interests
- Embodied AI
- Reinforcement Learning
- Foundation Model for Decision Making
Education
PhD in Computer Science, 2017-2022
Harbin Institute of Technology
Joint PhD Program, 2020-2022
University of Toronto
Publications
“✉” denotes corresponding author
Quickly discover relevant content by filtering publications.
Under Review. 2026
We introduce PRTS, a New Generation of Reinforcement Learning-Native Robotic Vision-Language-Action(VLA) foundation model.
In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) findings, 2026
Re^2MoGen combines LLM reasoning, keyframe-guided completion, and physics-aware RL refinement for open-vocabulary text-to-motion generation.
arXiv preprint
We introduce Pro-HOI, a generalizable framework for robust humanoid loco-manipulation via perceptive root-guided control.
In International Conference on Machine Learning (ICML), 2026
V2A unifies dynamics alignment, value alignment, and value assignment for heterogeneous cross-domain offline RL and improves robust source-data filtering.
arXiv preprint
Perception-Action integrated Decision-making (PAiD), a progressive framework for humanoid soccer skills.
In Robotics: Science and Systems (RSS), 2026
We address humanoid skateboarding, a highly challenging task requiring stable dynamic maneuvering on an humanoid platform via physics-aware whole-body control for humanoid skateboarding in dynamic settings.
In Robotics: Science and Systems (RSS), 2026
We introduce X-Loco, a framework for training a vision-based generalist humanoid locomotion controller through synergetic policy distillation.
In International Conference on Learning Representations (ICLR), 2026
We propose Align-Then-stEer (ATE), a framework that adapts VLAs to novel robots and tasks through unified latent guidance. ATE can handle significant domain shifts without compromising performance and compatible to Pi0, RDT, and etc.
arXiv preprint arXiv:2512.02834
We propose TACO, a test-time-scaling framework for VLAs that improves inference stability and success rates by preventing distribution shifts at test time.
In IEEE International Conference on Robotics & Automation(ICRA), 2026
We present VMS, a unified whole-body controller that enables humanoid robots to learn diverse and dynamic behaviors within a single policy through hybrid tracking and orthogonal mixture of experts.
In AAAI Conference on Artificial Intelligence (AAAI), 2026 Oral
We propose a framework for adaptive humanoid control via multi-behavior distillation and reinforced fine-tuning, achieving state-of-the-art performance and AAAI-26 Oral recommendation.
under review at ARR rounding 2026
We propose the Learn as Individuals, Evolve as a Team (LIET) framework to enable multi-agent LLMs to adapt to embodied environments through individual learning and team evolution
In IEEE International Conference on Multimedia & Expo (ICME), 2026
We propose a novel bimanual foundation policy that leverages text-to-video models to predict robot trajectories and uses optical flow as an intermediate variable to improve generalization.
In Pattern Recognition, 2025 (under review)
We propose a novel framework that enables humanoid robots to traverse complex terrains with controllable human-like gaits using a mixture of latent residual experts and multi-discriminators.
In Neural Information Processing Systems (NeurIPS), 2025
We propose a physics-based humanoid control framework, aiming to master highly-dynamic human behaviors such as Kungfu and dancing through multi-steps motion processing and adaptive motion tracking.
In Neural Information Processing Systems (NeurIPS), 2025
we propose HumanoidGen, an automated task creation and demonstration collection framework that leverages atomic dexterous operations and LLM reasoning to generate relational constraints.
In Neural Information Processing Systems (NeurIPS), 2025
We propose CURE, a method that splits LLM planning uncertainty into epistemic and intrinsic parts for more reliable robot decision-making.
In Neural Information Processing Systems (NeurIPS), 2025
We propose Diffusion-Inspired Multi-Agent world model (DIMA), a novel framework for multi-agent reinforcement learning that leverages diffusion models to reduce modeling complexity and improve sample efficiency.
In Neural Information Processing Systems (NeurIPS), 2025
We propose Adversarial Locomotion and Motion Imitation (ALMI) for humanoid robots, which serves as a novel framework for loco-manipulation tasks, enabling adversarial policy learning between upper and lower body.
In Neural Information Processing Systems (NeurIPS), 2025
We decompose the reward value in RLHF into two independent components that consists prompt-free reward and prompt-related reward, and propose a new reward learning algorithm by prioritizing data samples based on their prompt-free reward values.
In Journal of the American Statistical Association (JASA), 2025
This paper proposes an offline Wasserstein-based approach to estimate the joint distribution of multivariate discounted cumulative rewards, establishes finite sample error bounds in the batch setting, and demonstrates its superior performance through extensive numerical studies.
In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025
we propose a novel whole-body locomotion algorithm based on dynamic balance and Reinforcement Learning (RL) that enables humanoid robots to traverse extreme terrains, particularly narrow pathways and unexpected obstacles, using only proprioception.
In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025
we propose a novel Vision-Language Preference learning framework that learns a vision-language preference model to provide preference feedback for embodied manipulation tasks.
In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025
We develop a learning framework combining offline diffusion planner and online preference alignment with weak preference labeling for legged locomotion control.
In Annual Meeting of the Association for Computational Linguistics (ACL), 2025
We propose an online iterative self-alignment method for radiology report generation that iteratively generates unlimited preference data and automatically aligns with radiologists’ multiple objectives.
In Annual Meeting of the Association for Computational Linguistics (ACL), 2025
We propose a novel framework for multi-agent collaboration that introduces Reinforced Advantage feedback (ReAd) for efficient self-refinement of plans.
In International Conference on Machine Learning (ICML), 2025
We develop a versatile diffusion planner that can leverage large-scale inferior data that contains task-agnostic sub-optimal trajectories, with the ability to fast adapt to specific tasks.
In International Conference on Learning Representations (ICLR), 2025 Spotlight
We propose count-based online preference optimization for LLM alignment that leverages coin-flip counting to encourage exploration in online RLHF.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2025
We present a framework called Selective Myopic bEhavior Control~(SMEC), which results from the insight that the short-term behaviors of prior policies are sharable across tasks.
In International Conference on Learning Representations (ICLR), 2025
We propose a novel framework that generalizes demonstrations to establish critic-regularized grounding and optimization in the long-term planning of LLMs.
In International Conference on Learning Representations (ICLR), 2025
We introduce ExpoComm, a scalable communication protocol that leverages exponential topologies for efficient information dissemination among many agents in large-scale multi-agent reinforcement learning.
In Transactions on Machine Learning Research (TMLR), 2025
we propose a novel world model for MARL that learns decentralized local dynamics for scalability, combined with a centralized representation aggregation from all agents.
In AAAI Conference on Artificial Intelligence (AAAI), 2025
We propose a new radiology report generation method that aligns the pre-trained model with multiple human preferences via preference-guided multi-objective optimization reinforcement learning.
In AAAI Conference on Artificial Intelligence (AAAI), 2025
We propose Forward KL regularized Preference optimization for aligning Diffusion policies to align the diffusion policy with preferences, learning to align the policy output with human intents in various tasks.
In Artificial Intelligence (AIJ), 2025
This work designs and analyzes a novel set of algorithms for multi-agent reinforcement learning (MARL) based on the principle of information-directed sampling (IDS).
In Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2024
We introduce ODRL, the first benchmark tailored for evaluating off-dynamics RL methods where one needs to transfer policies across different domains with dynamics mismatch.
In Neural Information Processing Systems (NeurIPS), 2024
We adopt multi-task preferences as a unified condition for both single- and multi-task decision-making, and propose preference representations aligned with preference labels.
In Neural Information Processing Systems (NeurIPS), 2024
We introduce a novel framework that leverages a unified discrete diffusion to combine generative pre-training on human videos and policy fine-tuning on a small number of action-labeled robot videos.
In Annual Conference on Robot Learning (CORL), 2024 Oral
We propose a novel single-stage privileged knowledge distillation method called the Historical Information Bottleneck (HIB) to narrow the sim-to-real gap for legged locomotion.
In European Conference on Artificial Intelligence (ECAI), 2024
We propose a novel dynamic policy constraint that restricts the learned policy on the samples generated by the exponentional moving average of previously learned policies for offline RL.
In International Conference on Machine Learning (ICML), 2024
We propose a novel unsupervised RL framework via an ensemble of skills, where each skill performs partition exploration based on the state prototypes.
In International Conference on Machine Learning (ICML), 2024
We propose a novel representation-based approach to measure the domain gap, where the representation is learned through a contrastive objective by sampling transitions from different domains.
In International Conference on Machine Learning (ICML), 2024
We propose SAM-E, a novel architecture for robot manipulation by leveraging a vision-foundation model for generalizable scene understanding and sequence imitation for long-term action reasoning.
In International Conference on Machine Learning (ICML), 2024
We construct an example to show the information-theoretical improvement in sample efficiency achieved by goal relabeling and develop an RL algorithm called GOALIVE.
In International Conference on Machine Learning (ICML), 2024
We consider dynamics adaptation settings where there exists dynamics mismatch between the source domain and the target domain, and one can get access to sufficient source domain data, while can only have limited interactions with the target domain.
Neural Networks, 2024
We propose a novel Dynamic Skill Learning (DSL) framework to enable more effective adaptation and collaboration in complex tasks.
In IEEE International Conference on Robotics and Automation (ICRA), 2024 Oral
We consider a novel risk-sensitive perspective to enhance the robustness of legged locomotion.
In AAAI Conference on Artificial Intelligence (AAAI), 2024
We propose Optimistic Value Distribution Explorer (OVD-Explorer) to achieve a noise-aware optimistic exploration for continuous control.
In Journal of Artificial Intelligence Research (JAIR), 2023
We propose the Robust Offline-to-Online (RO2O) algorithm, designed to enhance offline policies through uncertainty and smoothness, and to mitigate the performance drop in online adaptation.
In Neural Information Processing Systems (NeurIPS), 2023
We aim to investigate the effectiveness of a single diffusion model in modeling large-scale multi-task offline data, which can be challenging due to diverse and multimodal data distribution.
In Neural Information Processing Systems (NeurIPS), 2023
We reveal the limitations of these methods and explore the problem from the value difference perspective via a novel insight on the value consistency across domains.
In SCIENCE CHINA Information Sciences, 2023
Our work builds upon the investigation of successor representations for task generalization in online RL and extends the framework to incorporate offline-to-online learning.
In Artificial Intelligence (AIJ), 2023
We propose an uncertainty-based MTDS approach that shares the entire dataset without data selection.
In International Conference on Machine Learning (ICML), 2023
We propose a novel unsupervised skill discovery method through contrastive learning among behaviors, which makes the agent produce similar behaviors for the same skill and diverse behaviors for different skills.
In Information Sciences, 2023
We introduce a novel strategy employing diverse randomized value functions to estimate the posterior distribution of Q-values.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023
We propose falSe COrrelation REduction (SCORE) for offline RL, a practically effective and theoretically provable algorithm.
In Neural Information Processing Systems (NeurIPS), 2022 Spotlight
We propose Robust Offline Reinforcement Learning (RORL) with a novel conservative smoothing technique.
IEEE Transactions on Systems, Man, and Cybernetics: Systems. 2022
We present an offline RL algorithm that combines hindsight relabeling and supervised regression to predict actions without oracle information.
In International Conference on Machine Learning (ICML), 2022 Spotlight
We study how RL can be empowered by contrastive learning in a class of Markov decision processes (MDPs) and Markov games (MGs) with low-rank transitions. For both models, we propose to extract the correct feature representations of the low-rank model by minimizing a contrastive loss.
IEEE Transactions on Neural Networks and Learning Systems, 2022
We propose monotonic quantile network (MQN) with conservative quantile regression (CQR) for risk-averse policy learning.
IEEE Transactions on Neural Networks and Learning Systems, 2022
We conduct a comprehensive survey on existing exploration methods for both single-agent RL and multiagent RL.
International Conference on Learning Representations (ICLR), 2022 Spotlight
We propose Pessimistic Bootstrapping for offline RL (PBRL), a purely uncertainty-driven offline algorithm without explicit policy constraints.
In Neural Information Processing Systems (NeurIPS), 2021
We propose a Dynamic Bottleneck (DB) model, which attains a dynamics-relevant representation based on the information-bottleneck principle.
In International Conference on Machine Learning (ICML), 2021 Spotlight
We propose a principled exploration method for DRL through Optimistic Bootstrapping and Backward Induction (OB2I).
IEEE Transactions on Cybernetics, 2021
We analyze the hindsight bias due to this use of hindsight goals and propose the bias-corrected HER (BHER), an efficient algorithm that corrects the hindsight bias in training.
IEEE Transactions on Neural Networks and Learning Systems, 2021
We propose a variational dynamic model based on the conditional variational inference to model the multimodality and stochasticity.
Knowledge-Based Systems (KBS), 2020
We propose a novel prioritized hindsight model for multi-goal RL in which the agent is provided with more valuable goals, as measured by the expected temporal-difference (TD) error.
Knowledge-Based Systems (KBS), 2020
This paper describes a method of obtaining more accurate estimated action values for CDRL using adaptive bounds.
Journal of Computer Research and Development (in Chinese), 2019
We propose an active sampling method based on TD-error adaptive correction in order to solve sample efficiency problem in deep Q-learning.
Talks
- 具身决策前沿：从强化学习到具身大模型PDF. 中国计算机学会. 2026
- 中国具身智能大会：迈向跨本体泛化的通用操作模型. 中国人工智能学会. 2026
- 光大证券年度投资策略会: 重构人类生产力,机器人点亮未来. 光大证券 2025.
- 半导体无线网络芯片技术论坛: 具身智能大小脑协同. 华为技术有限公司 2025.
- 清华大学深圳国际研究生院: SIGS专题讲座. 清华大学 2025.
- 人工智能研讨活动: 人工智能学会具身智能研讨会. 中国人工智能学会 2025.
- 中国空间智能大会: 空间智能与人形机器人论坛. 中国图象图形学学会 2025.
- 上海市计算机学会计算机视觉专委会学术年会: 视觉中的大模型学习论坛. 计算机学会 2025.
- 具身智能开发者大会: 人形机器人运动与交互专题论坛. 张江科学会堂 2025.
- VALSE Webinar 379: 具身智能中的多模态感知与精细操控. VALSE 2025.
- 中国具身智能大会: 模式识别与机器人论坛. 人工智能学会 2025.
- CSIG图像图形学科前沿讲习班: 小模型的曙光和机会之思辨. 图形图像学会 2025.
- 自主机器人技术研讨会（ARTS 2024）: 大模型驱动的具身智能. 自动化学会 2024.
- CCF决策智能会议（RLChina）: 具身智能体与机器人论坛. 计算机学会 2024.
- 中国数据挖掘会议: AI Agent与多智能体系统研讨会. 计算机学会 2024.
- 世界人工智能大会-WAIC: 青年优秀论文论坛. 上海科协 2024.
- 中国具身智能大会: 大模型与具身智能. 人工智能学会 2024.
- 上海自主智能无人系统科学中心: 智能决策学术报告. 同济大学 2023.
- 决策智能峰会: 决策大模型论坛. DataFun 2023.
- 多智能体强化学习讲习班: 离线强化学习和大模型. 自动化学会 2023.
- GAITC 全球人工智能技术大会: 人工智能原理专题论坛. 中国科协 2023.
- 高级人工智能公开课: 强化学习. 天津大学 2023.
- 从统计学到人工智能研讨会: 强化学习的不确定性估计. 上海财经大学 2022.
Service
- Senior Program Committee Member (SPC) / Area Chair (AC) of AAMAS (2024 - 2025) [proof]
- Area Chair (AC) of Pattern Recognition and Computer Vision (PRCV) (2025 - ) [proof]
- Industry Program Chair of IEEE International Conference on Multimedia and Expo (ICME) 2026 [proof]
- Program Committee Members (PC) / Conference Reviewer of RSS (2024 - 2025)
- Program Committee Members (PC) / Conference Reviewer of NeurIPS (2021 - 2025)
- Program Committee Members (PC) / Conference Reviewer of ICLR (2021 - 2025)
- Program Committee Members (PC) / Conference Reviewer of ICML (2022 - 2025)
- Program Committee Members (PC) / Conference Reviewer of AAAI (2021 - 2025)
- Program Committee Members (PC) / Conference Reviewer of ICRA (2024 - 2025)
- Program Committee Members (PC) / Conference Reviewer of ECAI (2023 - 2025)
- Journal Reviewer: IEEE Trans. PAMI, IEEE Trans. Cybernetics, IEEE Trans. TNNLS, IEEE Trans. TETCI, IEEE Trans. Intelligent Vehicles, Pattern Recognition.
Experience
Research Scientist, Lead of Embodied AI Team
TeleAI, China Telecom
Researcher
Shanghai AI Laboratory
Joint PhD Student
University of Toronto
PhD Student
Harbin Institute of Technology