Zewei Zhou

PhD Student @UCLA | Autonomous Agents, Self-Improvement

[Wenhao Ding @wenhaoding95]

Glad to share that our recent work, AlpaJudge for driving reasoning evaluation, was released at #GTCBerlin! Please try it on GitHub https://t.co/M7zFB4c4zT. Details of this model can be found in the VeriFine paper: https://t.co/wgEzq04w75

[Marco Pavone @drmapavone]

What if AI could improve not only its own policies, but also the judges that evaluate them?

As AI policies improve, they encounter new challenges and develop new failure modes. The judges evaluating them need to evolve as well. Otherwise, they can become a bottleneck to further progress. This is particularly relevant for #Physical #AI, where good decisions require spatial grounding, causal reasoning, and safety awareness.

In our new paper, we introduce VeriFine, an #agentic #harness for continuous self-improvement in Physical AI, where policies, training curricula, and judges evolve together.

The idea is to build agentic systems that can identify policy failures, turn them into useful training signals, and improve the policy. When the judge itself becomes a bottleneck, the harness brings in targeted human feedback to refine its evaluations and guide the next round of improvement.

We evaluate VeriFine on autonomous driving and robot navigation, demonstrating continued improvements in both policy and judge capabilities through reinforcement learning and supervised fine-tuning.

Key takeaway: We're developing agentic harnesses that go beyond optimizing a policy in a fixed loop. They adapt the improvement process itself—including how failures are identified, how training data is selected, and how progress is evaluated.

📄 Paper: https://t.co/NckfQWdqLQ
🌐 Website: https://t.co/8SSi8mFsPq

Work led by Zewei Zhou, in collaboration with Rachel Luo Yulong Cao @ChaoweiX Chensheng Peng @Boyiliee @tianran_ Zheng Lian @yan_wang_9 Jiaqi Ma @iamborisi @wenhaoding95

#PhysicalAI #Robotics #AgenticAI #Alpamayo

@NVIDIADRIVE @NVIDIAAI

[jiagengliu @jiagengliu02]

🚀 Introducing 𝐀𝐮𝐭𝐨𝐀𝐠𝐞𝐧𝐭𝟎: An 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐮𝐧𝐭𝐢𝐦𝐞 for Safe Closed-Loop Driving!

Most autonomous driving research focuses on optimizing E2E driving policies. However, in long-tail, 𝐥𝐚𝐫𝐠𝐞𝐥𝐲 𝐮𝐧𝐯𝐞𝐫𝐢𝐟𝐢𝐚𝐛𝐥𝐞 driving environments, stronger policies alone cannot ensure 𝐬𝐚𝐟𝐞 𝐚𝐧𝐝 𝐫𝐨𝐛𝐮𝐬𝐭 𝐜𝐥𝐨𝐬𝐞𝐝-𝐥𝐨𝐨𝐩 𝐞𝐱𝐞𝐜𝐮𝐭𝐢𝐨𝐧.

In 𝐀𝐮𝐭𝐨𝐀𝐠𝐞𝐧𝐭𝟎, we reformulate traditional runtime monitoring as an 𝐚𝐠𝐞𝐧𝐭𝐢𝐜 𝐫𝐮𝐧𝐭𝐢𝐦𝐞 𝐡𝐚𝐫𝐧𝐞𝐬𝐬 that reasons about 𝐰𝐡𝐞𝐧 𝐭𝐨 𝐢𝐧𝐭𝐞𝐫𝐯𝐞𝐧𝐞 and 𝐡𝐨𝐰 𝐭𝐨 𝐫𝐞𝐜𝐨𝐯𝐞𝐫, leveraging two key capabilities:

1️⃣ 𝐒𝐜𝐞𝐧𝐞 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐑𝐞𝐟𝐢𝐧𝐞𝐦𝐞𝐧𝐭 — Leveraging VLM reasoning to refine critical scene context missed by nominal policies, enabling reliable action verification and triggering interventions when nominal policies fail.

2️⃣ 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐞𝐜𝐨𝐯𝐞𝐫𝐲 𝐋𝐨𝐨𝐩 — Leveraging VLM-guided stochastic optimization to iteratively propose, verify, and refine explainable recovery policies through verifier feedback.

📊 𝐒𝐭𝐫𝐨𝐧𝐠 𝐜𝐥𝐨𝐬𝐞𝐝-𝐥𝐨𝐨𝐩 𝐠𝐚𝐢𝐧𝐬! 𝐀𝐮𝐭𝐨𝐀𝐠𝐞𝐧𝐭𝟎 consistently improves representative E2E driving policies on 𝐅𝐚𝐢𝐥𝟐𝐃𝐫𝐢𝐯𝐞 (up to +37.0 SR) and 𝐍𝐚𝐯𝐒𝐚𝐟𝐞-∞ (up to +32.8 SR), without policy retraining.

Our results highlight 𝐚𝐠𝐞𝐧𝐭𝐢𝐜 𝐫𝐮𝐧𝐭𝐢𝐦𝐞 𝐚𝐬 𝐚 𝐜𝐨𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐚𝐫𝐲 𝐝𝐢𝐫𝐞𝐜𝐭𝐢𝐨𝐧 𝐭𝐨 𝐬𝐜𝐚𝐥𝐢𝐧𝐠 𝐄𝟐𝐄 𝐝𝐫𝐢𝐯𝐢𝐧𝐠 𝐩𝐨𝐥𝐢𝐜𝐢𝐞𝐬.

🎉 Huge congrats to the amazing team: @jiagengliu02 , @sethzhao506 , @Zewei__Zhou , Alejandro Huerta, Aidan Bayer-Calvert, Yuxin Bao, @DarthUtopian , Robert Spataru, @zijun_cui02 , @yining_hong , @zhoubolei , and Jiaqi Ma!

🛠️ Everything is open-sourced!

🌐 Project: https://t.co/N7KMJZndiQ
⭐ Code: https://t.co/VXTVXlNeRA
📄 Paper: https://t.co/AdbYhXU2Em

#AgenticAI #AutonomousDriving #EndToEndDriving #PhysicalAI

[Zewei Zhou @Zewei__Zhou]

@wenhaoding95 Thanks for your mentorship and support, too! 💚

[Zewei Zhou @Zewei__Zhou]

@nvidia @drmapavone @iamborisi @wenhaoding95 Huge thanks to all my co-authors for their contributions and collaboration 🙌:
Rachel Luo, Yulong Cao, @ChaoweiX, Chenshen Peng, @Boyiliee, @tianran_ , Zheng Lian, @yan_wang_9, Jiaqi Ma, @iamborisi, @drmapavone, @wenhaoding95.

[Zewei Zhou @Zewei__Zhou]

🤔 𝘾𝙖𝙣 𝙨𝙚𝙡𝙛-𝙞𝙢𝙥𝙧𝙤𝙫𝙚𝙢𝙚𝙣𝙩 𝙗𝙚 𝙨𝙪𝙨𝙩𝙖𝙞𝙣𝙚𝙙 𝙬𝙞𝙩𝙝 𝙖 𝙛𝙞𝙭𝙚𝙙, 𝙣𝙤𝙣-𝙘𝙤-𝙚𝙫𝙤𝙡𝙫𝙞𝙣𝙜 𝙟𝙪𝙙𝙜𝙚?

Better policies expose new failures. If their judges stay fixed, verification can become the bottleneck to further self-improvement.

[Zewei Zhou @Zewei__Zhou]

This work was done during my internship at @nvidia . I’m grateful to @drmapavone and @iamborisi for their guidance and support. A special thank you to my mentor, @wenhaoding95, for his mentorship throughout this project.

[Qualcomm Research & Technologies @QCOMResearch]

Congrats to the Qualcomm Innovation Fellowship NA winners 👏
PhD research + mentorship from @Qualcomm engineers = real-world impact.
See what they’re building:
https://t.co/I5e1ONUyNk

[Zewei Zhou @Zewei__Zhou]

Huge thanks to my amazing co-authors and advisors:
Tianhui Cai, @sethzhao506, YUN ZHANG,
Dr. @ZhiyuChris, Prof. Dr. @zhoubolei, Prof. Dr. Jiaqi Ma

[Zewei Zhou @Zewei__Zhou]

Thrilled to share that our paper 𝐀𝐮𝐭𝐨𝐕𝐋𝐀 has been accepted to 𝐍𝐞𝐮𝐫𝐈𝐏𝐒 𝟐𝟎𝟐𝟓! See you in San Diego! 🦭

🗓️ Dec. 3 (Wed), 11:00 AM – 2:00 PM PST
📍 Exhibit Hall C,D,E # 2208

[Zewei Zhou @Zewei__Zhou]

📎 𝘋𝘪𝘷𝘦 𝘋𝘦𝘦𝘱𝘦𝘳 𝘪𝘯𝘵𝘰 𝘈𝘶𝘵𝘰𝘝𝘓𝘈:
𝐏𝐫𝐨𝐣𝐞𝐜𝐭 𝐰𝐞𝐛𝐬𝐢𝐭𝐞: https://t.co/f1C5t8KM28
𝐏𝐚𝐩𝐞𝐫 𝐥𝐢𝐧𝐤: https://t.co/Um22mjIkwG

[Zewei Zhou @Zewei__Zhou]

@sethzhao506 @ZhiyuChris @zhoubolei @Ross__Greer @rruisong @AlinaRoitberg Jiajie (Vincent) Zhang, Dr. Julie Stephany Berrio Perez, Chuheng Wei, @walk2out, Prof. Dr. @_vztu, Prof. Dr. @nightRome, Prof. Dr. @DanielWatzenig, Prof. Dr. Mohan Trivedi, Prof. Dr. Alois Knoll

[Zewei Zhou @Zewei__Zhou]

🏅 𝑩𝒆𝒔𝒕 𝑷𝒂𝒑𝒆𝒓 𝑨𝒘𝒂𝒓𝒅 (𝟏𝐬𝐭) @ ICCV 2025 DriveX Workshop!

"𝐕𝟐𝐗𝐏𝐧𝐏: 𝐕𝐞𝐡𝐢𝐜𝐥𝐞-𝐭𝐨-𝐄𝐯𝐞𝐫𝐲𝐭𝐡𝐢𝐧𝐠 𝐒𝐩𝐚𝐭𝐢𝐨-𝐓𝐞𝐦𝐩𝐨𝐫𝐚𝐥 𝐅𝐮𝐬𝐢𝐨𝐧 𝐟𝐨𝐫 𝐌𝐮𝐥𝐭𝐢-𝐀𝐠𝐞𝐧𝐭 𝐏𝐞𝐫𝐜𝐞𝐩𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧"

🚘 https://t.co/GWuzgPcMSI

[Zewei Zhou @Zewei__Zhou]

@sethzhao506 @ZhiyuChris @zhoubolei Sincerely, thanks to the workshop organizers for making this wonderful workshop happen:
Dr. Walter Zimmer, Prof. Dr. @Ross__Greer, Dr. @rruisong, Dr. Arpita V., Dr. Lars Ullrich, Max Ronecker, Xingcheng Zhou, Yifan Liu, Haoxuan Ma, Rahul Raja, Prof. Dr. @AlinaRoitberg