Zike Yan (颜子轲)

I am currently a postdoctoral fellow at Institute for AI Industry Research (AIR), Tsinghua University, working with Prof. Guyue Zhou and Prof. Ya-Qin Zhang. Prior to that, I obtained my Ph.D Degree from Peking University in 2023, advised by Prof. Hongbin Zha.

My research is driven by the long-term goal of developing perceptually and physically adaptive intelligence for mobile agents deployed over extended periods. I believe that active learning through live interactions with the environment is the key to scaling such intelligence with unending streaming data, provided that the system can adapt safely and efficiently. I focus on vision-based adaptive systems through continual learning, mainly for dense reconstruction and exploration, with recent extensions to robotic manipulation and locomotion. My primary interest is balancing continual adaptation and knowledge retention. This involves developing suitable representations and optimization strategies to meet the real-time constraints of diverse robotic tasks. A summary of my previous work can be found in the Research Summary and Publications.

Contacts: Twitter Github Google scholar Email

News

[June 2025] Two papers accepted to ICCV 2025!
[June 2025] DISCOVERSE accepted to IROS 2025!
[May 2025] ActiveSplat accepted to RA-L!
[May 2025] I was recognized as a CVPR2025 Outstanding Reviewer!
[May 2025] I served as an Associate Editor at IROS 2025.
[Jul 2024] Two papers accepted to ECCV 2024!
[Jul 2024] Two papers accepted to IROS 2024!
[May 2024] I served as an Associate Editor at IROS 2024.
[Feb 2024] One paper accepted to CVPR 2024!

Research Summary

In the short term, my target systems can be summarized into three projects, which constitute the core focus of most of my publications. I envision these projects as three progressive phases toward achieving adaptive intelligence in robotic navigation, manipulation, and locomotion. The first phase, namely LiveMimic, seeks human-in-the-loop policy training to accelerate the pre-training stage with convenient in-distribution human demonstrations; The second phase aims to improve the policy generalization through realistic simulation (or a learned world model that predicts future state conditioned on action commands), namely ByteYard, thereby narrowing sim-to-real gap with safe post-training via trial-and-error; The final stage, namely Xplore, aims to promote autonomous and active refinement of deployed policies for test-time adaptation. My personal interest lies in studying the exploration-exploitation tradeoff within the neural network energy landscape, focusing on appropriate representations and corresponding optimization strategies to balance exploration efficiency and knowledge coverage in practice.

Online dense reconstruction

What is the right representation for compressing the streaming data?

How to achieve a balance between efficient adaptation and long-lasting memorization?

Show/Hide Work on Online Dense Reconstruction

Representations
- Flexible Gaussian primitives:
  [CVPR2021][RA-L2025]
- Compact neural representations:
  [ICCV2021][ICCV2023]
- + Object-decomposed structure:
  [ECCV2024][IROS2025]
Optimization Strategies
- Online Bayesian learning:
  [CVPR2021]
- Experience replay:
  [ICCV2021]
- + Generalization-forgetting trade-offs:
  [ICCV2023] [IROS2024]
- + Learn to memorize and to forget:
  [ECCV2024]

Active perception and learning for robotic applications

How can the robot efficiently learn and adapt the skills from natural supervisions?

How can the digitalized 3D information be exploited for goal-oriented task execution?

Show/Hide Work on Adaptive Robotic Systems

State Estimation
- Dynamic SLAM:
  [ECCV2024]
- + Object decomposition and pose estimation:
  [ICCV2025]
Autonomous Exploration
- From the optimization perspective:
  [ICCV2023]
- + Graph-based planning:
  [IROS2024]
- + Efficient Gaussian splatting & hierarchical planning:
  [RA-L2025]
Manipulation
- Realistic simulation for manipulation:
  [IROS2025]
- + Real2sim2real via differentiable simulation : [Arxiv2025]

Selected Publications ^{[Full list]}

ICCV 2025 ^{[Project: LiveMimic} ^ByteYard]

Proactive Scene Decomposition and Reconstruction

Baicheng Li, Zike Yan^†, Dong Wu, Hongbin Zha^†

IROS 2025 ^{[Project: ByteYard]}

DISCOVERSE: Efficient Robot Simulation in Complex High-Fidelity Environments ^{[Project page]} ^[Paper]

Many authors

IEEE RA-L 2025 ^{[Project: Xplore]}

ActiveSplat: High-Fidelity Scene Reconstruction through Active Gaussian Splatting ^{[Project page]} ^[Paper]

Yuetao Li^*, Zijia Kuang^*, Ting Li, Zike Yan^†, Guyue Zhou, Shaohui Zhang^†

ECCV 2024 ^{[Project: Xplore]}

Learn to Memorize and to Forget: A Continual Learning Perspective of Dynamic SLAM ^[Paper]

Baicheng Li, Zike Yan^†, Dong Wu, Hanqing Jiang, Hongbin Zha^†

ECCV 2024 ^{[Project: ByteYard]}

Structured-NeRF: Hierarchical Scene Graph with Neural Representation ^[Paper]

Zhide Zhong^*, Jiakai Cao^*, Songen Gu, Sirui Xie, Liyi Luo, Hao Zhao, Guyue Zhou, Haoang Li, Zike Yan^†

IROS 2024 ^{[Project: Xplore]}

Active Neural Mapping at Scale ^{[Project page]} ^[Paper]

Zijia Kuang, Zike Yan^†, Hao Zhao, Guyue Zhou, Hongbin Zha

CVPR 2024 ^{[Project: ByteYard]}

PanoRecon: Real-Time Panoptic 3D Reconstruction from Monocular Video ^[Paper]

Dong Wu, Zike Yan, Hongbin Zha

ICCV 2023 ^{[Project: Xplore]}

Active Neural Mapping ^{[Project page]} ^[Paper]

Zike Yan, Haoxiang Yang, Hongbin Zha

CICAI 2023 ^{[Best Paper Runner-up,}^{Project: ByteYard]}

MARS: An Instance-aware, Modular and Realistic Simulator for Autonomous Driving ^{[Project page]} ^[Paper]

Many authors

IEEE RA-L 2022 ^{[Project: LiveMimic]}

Visual-Inertial Odometry based on Kinematic Constraints in IMU Frames ^[Paper]

Xin Wang, Youqi Pan, Zike Yan, Hongbin Zha

ICCV 2021 ^{[Project: Xplore]}

Continual Neural Mapping: Learning an Implicit Scene Representation from Sequential Observations ^{[Project page]} ^[Paper]

Zike Yan, Yuxin Tian, Xuesong Shi, Ping Guo, Peng Wang, Hongbin Zha

IEEE T-RO 2021 ^{[Project: LiveMimic]}

Line Flow based Simultaneous Localization and Mapping ^[Paper]

Qiuyuan Wang, Zike Yan, Junqiu Wang, Fei Xue, Wei Ma, Hongbin Zha

CVPR 2021 ^{[Project: Xplore]}

Online Learning of a Probabilistic and Adaptive Scene Representation ^[Paper]

Zike Yan, Xin Wang, Hongbin Zha

Virtual Reality and Intelligent Hardware 2019

Flow-based SLAM: From Geometry Computation to Learning ^[Paper]

Zike Yan, Hongbin Zha

About

Services

Reviewer: CVPR, ICCV, ECCV, ICLR, NeurIPS, ICRA, IROS, CoRL, 3DV, ACMMM, IJCV, T-VCG, T-IV, RA-L.
Associate Editor: IROS2024, IROS2025.

Other Things

My given name, Zike (子轲 /zɪkə/), is inspired by two great Chinese philosophers: Confucius and Mencius. I am the 84th generation of the Yan family (颜 /yɛn/), which traces its origins to Yan Hui, a renowned disciple of Confucius.
I am a petrolhead and a wildlifer. I enjoy solo trips (during the conferences).
I am looking for potential collaborations in areas where I have limited expertise, including: lifelong/generalizable robot learning, the emergence phenomenon in learned networks, continual learning as computationally-constrained RL, and exploration for LLMs. Feel free to reach out!