01 Training · Action Modeling
NeurIPS / 2025
Wenbo Zhang.
Ph.D. Candidate · Embodied AI · Robotic Manipulation
Graduating Looking for opportunities in agentic robotics.
Now
Visiting Student with Mengdi Xu @ Tsinghua IIISResearch Intern with Xiao Ma @ ByteDance Seed, BeijingPhD Candidate @ Adelaide University, Australia
New · Sep 2026
An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond
Co-first author & project lead · RoboDojo Team
ReportarXivCode
Research
Building physical agents?Why not virtual agents? Sadly, I believe that space is — and will increasingly be — rapidly converging under the digital hegemony of a few big companies. that self-improve through experience.
I work on embodied AI and robotic manipulation — toward autonomous systems that learn from their own experience and help with everyday tasks, without a human in the loop.
The Ideal — an embodied analogue of the three-stage recipe behind GPT
01
Native Multimodal Pretraining
≈ pretraining · compress knowledge
→
02
Policy Extraction
≈ post-training · extract knowledge
→
03
Align with Real-World
≈ alignment · to the final goal
A journey of a thousand miles begins with a single step | 千里之行，始于足下
01 / Training
Action Modeling
Chain-of-Action
→
02 / Post-Training
SFT Strategies
Effective Tuning
→
03 / Deployment
Test-Time Training
TTT-VLA
↻ Data Flywheel
↻ closes the data flywheel → training
02 Post-Training · SFT Strategies
ICRA / 2025
Effective Tuning Strategies for Generalist Robot Manipulation Policies
03 Deployment · Test-Time Training
Preprint / 2026
TTT-VLA: Test-Time Latent Prompt Optimization for Vision-Language-Action Models
Profile / CV
News
- Sep 2026 An Unexpected Robot Policy is out: with no robot training, GPT-6 Astra acts through a fixed harness (L3) and ranks #1 on RoboDojo-Sim, above every trained VLA. Report arXiv
- Sep 2025 Chain-of-Action accepted to NeurIPS 2025 — trajectory autoregressive modeling for robotic manipulation. Website Code
- Jan 2025 Effective Tuning Strategies for Generalist Robot Manipulation Policies accepted to ICRA 2025. arXiv
- Nov 2024 Started a research internship at ByteDance Seed (Beijing), working on embodied AI, mentored by Xiao Ma.
- Feb 2024 Ranked 1st at the Robotic Vision Summer School (RVSS) Competition, Australia. Details
- Jan 2024 MMT-Bench accepted to ICLR 2024 — a comprehensive multimodal benchmark for large vision-language models. arXiv
- Mar 2023 Began my Ph.D. at the Australian Institute for Machine Learning (AIML), University of Adelaide.
Experience
Research Intern · Beijing, CN · mentor: Xiao Ma
Ph.D. Candidate · Adelaide, AU
Shanghai AI Laboratory
Research Intern · Shanghai, CN
SenseTime Research
Research Intern · CN
Sichuan University
M.Eng., Computer Science · Chengdu, CN
Service & Talks
ICCV VCL Challenges
Visual Continual Learning challenge
Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation
Lumina Embodied AI & RL China
Student Representative Speech, Commencement Ceremony
Graduate School of Computer Science, Sichuan University
CONFERENCE / JOURNAL REVIEWER
Awards
Adelaide Research Scholarship
University of Adelaide
China National Scholarship
Ministry of Education, China
China Graduate AI Innovation Competition — Huawei Special Award & 2nd Prize
Ministry of Education, China
Top 100 Students
Sichuan University
Outstanding Graduation Thesis
Sichuan University
