Dichotomous Diffusion Policy Optimization
Positive policy + Negative policy + CFG = Controllable inference policy
Zh1hao Wang
Reality · Robotics · Reinforcement · Rewards
Currently conducting research at ByteDance Seed Robotics.
A second-year master's student at School of Advanced Manufacturing and Robotics, Peking University, advised by Professor Junzhi Yu.
Also a research intern at the Institute for AI Research (AIR), Tsinghua University, advised by Professor Xianyuan Zhan.
Prior to that, I earned my B.Eng. degree in Intelligent Automotive Engineering from the Harbin Institute of Technology (HIT).
My research interests mainly focus on building Reliable and Scalable robot learning systems.
News
- 01/2026: X-VLA and DIPOLE have been accepted by ICLR 2026!
- 10/2025: X-VLA won the 🏆 Championship at AgiBot World Challenge @ IROS 2025!
- 05/2025: Our paper LBP (Latent space Backward Planning) has been accepted by ICML 2025!
- 04/2025: Our paper Novel ViDAR with RL-based Active SLAM has been accepted by TII 2025!
- 02/2025: Our paper UniAct has been accepted by CVPR 2025!
- 01/2025: Our paper Robo_MUTUAL has been accepted by ICRA 2025!
Show more ↓
Thanks to all collaborators and supporters, looking forward to more exciting research ahead!
Experience
Research
My long-term vision is to build the BRIDGE between artificial intelligence and the physical world — shaping truly embodied intelligence. 🤖🧠
X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
A highly scalable cross-embodiment VLA model that achieves remarkable performance with a tiny model size.
PhysiAgent: An Embodied Agent Framework in Physical World
An autonomous scaffolding framework to seamlessly integrate VLA and VLM into real-world embodied agents.
Efficient Robotic Policy Learning via Latent Space Backward Planning
A lightweight visuomotor policy controls robots with latent, backward, recursive subgoals.
A Novel ViDAR Device With Visual Inertial Encoder Odometry and RL-Based Active SLAM Method
A novel ViDAR device with reinforcement learning-based active SLAM method.
UniAct: Universal Actions for Enhanced Embodied Foundation Models
A new embodied foundation modeling framework operating in the Universal Action Space.
Robo-MUTUAL: Robotic Multimodal Task Specification via Unimodal Learning
An approach for training a multimodal robotic policy via only unimodal datasets.
* Equal contribution.
Awards & Achievements
- [2025] Peking University Graduate Student Special Academic Scholarship
- [2024] 1st at AIR Summer 2024 Projects Competition
- [2024] Shandong Outstanding Graduates
- [2023] Undergraduate National Scholarship
- [2023] Huawei Intelligent Foundation Scholarship
- [2022] 1st at Formula Student Autonomous China Competition
Academic Service
- Reviewer at conferences: AAAI 2026, CoRL 2026, ICLR 2026, NeurIPS 2026
- Reviewer at workshops: EWM@NeurIPS 2025, WRL@ICLR 2025
Misc
Outside of research, I also enjoy:
- 🎬 Movies: My favorite genres are sci-fi, horror, and thriller. My go-to theater is Beijing Zhongguancun Perfect World & IMAX.
- 🎵 Music: I mostly listen to R&B and K-POP. My favorite artists are Khalil Fong (方大同), 宇多田ヒカル (Hikaru Utada), EXO, and BTS.
- 🚴 Cycling: My current ride is a CAMP RADON, my daily commuter and weekend companion around Beijing.
- 🐱 Cat: I'd love to have a kitten of my own someday!