Bolei Zhou ✈️ CoRL'26 (@zhoubolei) header
Bolei Zhou ✈️ CoRL'26 (@zhoubolei) profile picture

Associate Professor at the Computer Science Department @UCLAComSci @UCLAengineering @UCLA and Chief AI Scientist at @CocoRobotics. Opinions are my own.

Los Angeles
Joined October 2010
Photo by Bolei Zhou ✈️ CoRL'26 (@zhoubolei)Photo by Bolei Zhou ✈️ CoRL'26 (@zhoubolei)Photo by Bolei Zhou ✈️ CoRL'26 (@zhoubolei)Photo by Bolei Zhou ✈️ CoRL'26 (@zhoubolei)Photo by Bolei Zhou ✈️ CoRL'26 (@zhoubolei)Photo by Bolei Zhou ✈️ CoRL'26 (@zhoubolei)Photo by Bolei Zhou ✈️ CoRL'26 (@zhoubolei)
🤖 Excited to share Cue the Flow, our paper to be presented at #CoRL2026 @corl_conf, led by Haoxuan. We steer frozen flow-matching VLAs with lightweight spatial cues for open-world delivery manipulation, achieving up to 2× task success with negligible overhead! Looking forward to seeing everyone in Austin tomorrow! 🔗 hatchetproject.github.io/del…
5
6
1
42
2,823
🚀 Excited that SidewalkBench will be presented as a Spotlight Paper at the upcoming @corl_conf #CoRL2026 by Zhizheng! 🤖 A scalable simulation benchmark for safe sidewalk robot navigation, featuring diverse urban environments and dynamic pedestrians. A great collaboration with @CocoRobotics. Thanks to the NVIDIA Academic Grant Program #NVIDIAGrant and @NVIDIAAIDev for hardware support. 💻 Code is available at vail.cs.ucla.edu/SidewalkBen…
5
23
1,949
🚀 Excited to share FlowPilot, our latest work on preference-based learning for autonomous sidewalk navigation, to be presented at CoRL'26 @corl_conf Instead of relying solely on expert demonstrations, FlowPilot learns from human preferences to develop navigation policies that better align with human expectations in complex urban environments. 🤖 Looking forward to the presentation and discussion in Austin next week! Website: vail.cs.ucla.edu/FlowPilot/. Work led by incredible Honglin He, and thanks to @CocoRobotics for the robot hardware and data support!
2
1
34
2,805
Bolei Zhou ✈️ CoRL'26 retweeted
A wheelchair that learns how its rider wants to move. That's what makes ChairNav interesting. Researchers at UCLA developed ChairNav, an autonomous wheelchair navigation system that learns from existing robot data, human driving demonstrations, and rider corrections. The training process has three stages: - Pre-training: Adapt camera observations and motion data from delivery robots. - Mid-training: Learn wheelchair navigation from human driving demonstrations. - Post-training: Convert rider interventions into preference feedback so the system improves its driving behaviour. It also tackles long-horizon navigation in real environments, including obstacles, narrow passages and outdoor routes. A promising direction for assistive robotics, where autonomy must adapt to the person using it.
♿ Excited to introduce ChairNav, our latest research toward long-horizon autonomous wheelchair navigation! 🤖 We explore how vision-based navigation foundation models can enable intelligent navigation in complex real-world sidewalk environments. A step toward physical AI that truly empowers people! Project & demos: vail.cs.ucla.edu/ChairNav/
1
5
838
♿ Excited to introduce ChairNav, our latest research toward long-horizon autonomous wheelchair navigation! 🤖 We explore how vision-based navigation foundation models can enable intelligent navigation in complex real-world sidewalk environments. A step toward physical AI that truly empowers people! Project & demos: vail.cs.ucla.edu/ChairNav/
4
5
2
62
4,515
Bolei Zhou ✈️ CoRL'26 retweeted
🚀 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: jiagengliu02.github.io/AutoA… ⭐ Code: github.com/jiagengliu02/Auto… 📄 Paper: arxiv.org/abs/2610.18866 #AgenticAI #AutonomousDriving #EndToEndDriving #PhysicalAI
1
8
1
14
1,454
Bolei Zhou ✈️ CoRL'26 retweeted
Introducing NavSafe-∞: an open, photorealistic CL benchmark with 280 event-based scenarios across 28 safety-critical event types. Across 20 E2E policies, we expose the OL-to-CL safety gap. Project, paper & code: navsafe-vail.github.io (1/5)
8
9
1
37
1,123
I took this video at IROS’26 in Pittsburgh two days ago. At first glance, it looks like an impressive autonomous humanoid making cocktail drinks. Zoom out, and you see what is really behind it. I have mixed feelings about watching this. 1. Teleoperation is getting remarkably good. Data collection is becoming more accurate, scalable, and immersive. Combined with rapidly improving AI models, I can imagine much of this eventually becoming autonomous. That is exciting, but it also makes the potential impact on human jobs feel much more immediate. 2. But what are we actually seeing in a robot demo? Without proper disclosure or reference, it can be surprisingly difficult to distinguish teleoperation from autonomy. As physical AI attracts more attention and investment, demos need to be much clearer about how much is autonomous and how much is human-controlled. 3. Even “autonomous” demos can be difficult to interpret. Is the robot actually generalizing? Or is it replaying a demonstration, following a scripted trajectory, or overfitting to one carefully prepared environment? A tabletop demo alone often can’t tell us. Maybe the important question for Physical AI is no longer just “Can the robot do it?” It is also: “How did the robot do it, and will it still work when we change the task, objects, or environment?”
9
20
8
103
45,256
I’ll be speaking at #IROS2026 in Pittsburgh this Sunday (Sept. 27) at two workshops! 🤖 Bridging Navigation — 8:45–9:30 AM 🤖 Social Robot Navigation — 11:40–11:55 AM I’ll share our work on scaling long-horizon sidewalk navigation — from learning from fleet data to world models and cross-embodiment. If you’re at IROS, come say hi! Always happy to chat and meet in person.
3
4
39
2,494
Bolei Zhou ✈️ CoRL'26 retweeted
We're ranked the #1 public university in the nation! This is a win for every Bruin who makes UCLA extraordinary. We couldn’t have done it without you. 💙💛 ucla.in/4he2HL2 #1UCLA
63
222
64
1,028
139,566
My first ICLR paper was in 2015: Object Detectors Emerge in Deep Scene CNNs (arxiv.org/pdf/1412.6856). That year, ICLR received 143 submissions and accepted 31 papers; ours was one of 12 oral papers. Fast forward to ICLR 2027: 60,000 abstract submissions. What an incredible journey for ICLR and for the field of AI.
5
2
73
11,850
Bolei Zhou ✈️ CoRL'26 retweeted
Coco AI interns really learned a lot this summer 🥲
2
1
12
20,186
I’ll be speaking at the Agentic AI Summit @ Berkeley tomorrow, where I’ll share our latest findings on scaling sidewalk navigation with world simulation at @UCLAComSci and @CocoRobotics . If you’re attending, welcome to join the Robotics & World Models Atlas session tomorrow afternoon. I’d be also happy to chat in person! rdi.berkeley.edu/events/agen…
4
11
1
84
8,698
Augmenting multi-agent world models with world state registers! Nice work led by Sicheng and Yuheng.
If multiple agents live in the same world, why should their video streams live in separate memories? Let’s make every player remember the same world state. Introducing W^2: Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers Project Page: vail-ucla.github.io/worldwea… [1/5]
2
29
5,904