Dai-Jie Wu吳岱桀
Ph.D. student · SPARK Lab, Kahlert School of Computing, University of Utah
Affiliated with the NSF Center for Human and Robot Co-Adaptation (HRC)
Seeking a Summer 2027 research internship
I am a first-year Ph.D. student advised by Prof. Kenneth Marino. Previously, I interned at SHENNONGSHI.AI, focusing on robot learning for cooking applications, and also collaborated with NTU Robot Learning Lab, advised by Prof. Shao-Hua Sun.
Research
I study how robot foundation models can adapt to new tasks, environments and robots, learn from data that scales, and work safely around people.
Adapting at deploymentFast adaptation to new tasks, environments and robots, without retraining.
Robots, like phones and computers, will keep getting cheaper and more varied, and a policy that only works where and on what it was trained cannot keep up. I want models that adapt quickly at deployment to a new task, a new environment or a new robot, through test-time adaptation and in-context learning rather than retraining. For new robots, I want to measure where today's models fail to transfer, starting with benchmarks built for that question.
Related: READ (adapting from five demonstrations without test-time gradient updates), CooT (in-context adaptation to unfamiliar partners).
Data that scalesBeyond teleoperation: simulation built by agents, and human data, from videos to handheld grippers.
Teleoperation alone will not give robots enough data. Two sources scale further: simulation, which agents can now build from real videos and images and use for data, benchmarks and evaluation; and human data, from videos of people to handheld UMI-style demonstrations of contact-rich skills like twisting off a bottle cap. Neither is tied to one robot, and I want to combine them.
Related: DIFO (imitation from observation, without action labels), LCIL (learning from suboptimal demonstrations), Intervention-Robust VLA (diverse demonstrations for deployment).
Around peopleInteraction and safety for robots that share space with people, starting in hospitals.
Most robot problems assume a robot alone in a room. Through the NSF HRC, our Utah team will deploy robots at the Craig H. Neilsen Rehabilitation Hospital, where a robot has to coordinate with people, adapt to them and be steered by them, and where safety goes beyond collisions to what its actions cause around people.
Related: SAD-Flower and ASAP (generative planning that stays within safety constraints), Intervention-Robust VLA (policies that keep working under a safety filter).
I use VLAs, world action models and agentic frameworks where each is strongest; agentic frameworks are too slow for low-level control, but can help build simulations and collect data. I care about evaluation that explains why methods fail.
News
- Nov 2026Released open, community-editable lists of CoRL 2026 and RSS 2026 accepted papers with project pages, code and abstracts; pull requests welcome (CoRL repo · RSS repo).
- Oct 2026READ is accepted to two CoRL 2026 workshops, Pretrain to Adapt and RoboMem.
- Oct 2026Language-Critique Imitation Learning is accepted to two CoRL 2026 workshops, Oopsie and LfC.
- Oct 2026Selected as a Top Reviewer for NeurIPS 2026 (top 8%).
- Aug 2026Our Utah team joins the new NSF Center for Human and Robot Co-Adaptation, led by UT Austin, to study how robots and people learn to live and work together. UT News · @theU
- May 2026Two papers, SAD-Flower and CooT, are accepted to ICML 2026, and I was selected as a Gold Reviewer. See you in Seoul!
- Sep 2025Will be attending CoRL and Humanoids 2025 conferences. See you in Seoul!
- Aug 2025Gave a talk on robot learning tutorial at ROS CON TAIWAN 2025! Slides
- May 2025Received PhD fellowship from Kahlert School of Computing at University of Utah!
- Apr 2025Our CRAB benchmark has been accepted to ACL Findings 2025!
- Oct 2024Our Diffusion Imitation from Observation paper has been accepted to NeurIPS 2024!
Publications
Adapting at deployment
-
READ: What Makes a Pretrained VLA Teachable by Demonstration? Teaches a pretrained VLA a new task from five demos, with no gradient steps. Dai-Jie Wu, Minh Pham Dinh, Kenneth Marino CoRL 2026 Workshops · Pretrain to Adapt, RoboMem
-
CooT: Learning to Coordinate In-Context with Coordination Transformers Adapts to unfamiliar partners in context, with no parameter updates. Huai-Chih Wang, Hsiang-Chun Chuang, Hsi-Chun Cheng, Dai-Jie Wu, Shao-Hua Sun ICML 2026 · ICML 2025 MAS WorkshoparXiv
Data that scales
-
Language-Critique Imitation Learning from Suboptimal Demonstrations Learns from imperfect demos through language critiques, not a single score. Chih-Han Yang, Dai-Jie Wu*, Yun-Ping Huang*, Ping-Chun Hsieh, Kenneth Marino, Shao-Hua Sun CoRL 2026 Workshops · Oopsie, LfCICLR 2027 under review
-
Diffusion Imitation from Observation Imitation from state-only demos, with a diffusion model as the discriminator. Bo-Ruei Huang, Chun-Kai Yang, Chun-Mao Lai, Dai-Jie Wu, Shao-Hua Sun NeurIPS 2024arXivCode
Around people
-
ASAP: Arbitrary-Step Safe Planning with MeanFlow and Control Barrier Functions Safe generative planning in as few as one step, even under unseen constraints. Tzu-Yuan Huang, Yi Yue, Dai-Jie Wu, Stefan Sosnowski, Shih-Hsin Wang, Kenneth Marino, Armin Lederer, Sandra Hirche ICLR 2027under review -
Intervention-Robust VLA: Diverse Demonstrations for Robot Deployment with Safety Filters Diverse demos keep a VLA working when a safety filter steps in. Tzu-Yuan Huang, Shao-Hsuan Chang, Dai-Jie Wu, Ralf Römer, Yung-Hui Li, Stefan Sosnowski, Angela P. Schoellig, Kenneth Marino, Sandra Hirche ICRA 2027under review -
SAD-Flower: Flow Matching for Safe, Admissible, and Dynamically Consistent Planning Flow-matching planning that guarantees safety, action limits and dynamics. Tzu-Yuan Huang, Armin Lederer, Dai-Jie Wu, Xiaobing Dai, Sihua Zhang, Hsiu-Chin Lin, Shao-Hua Sun, Stefan Sosnowski, Sandra Hirche ICML 2026arXiv
Other work
-
CRAB: Cross-environment Agent Benchmark for Multimodal Language Model Agents A benchmark for multimodal agents whose tasks span desktop and phone. Tianqi Xu*, Linyao Chen*, Dai-Jie Wu*, Yanjun Chen*, Zecheng Zhang, Xiang Yao, Zhiqiang Xie, Yongchao Chen, Shilong Liu, Bochen Qian, Anjie Yang, Zhaoxuan Jin, Jianbo Deng, Philip Torr, Bernard Ghanem, Guohao Li ACL Findings 2025 · NeurIPS 2024 OWA WorkshoparXivCode
-
READ: What Makes a Pretrained VLA Teachable by Demonstration? Teaches a pretrained VLA a new task from five demos, with no gradient steps. Dai-Jie Wu, Minh Pham Dinh, Kenneth Marino CoRL 2026 Workshops · Pretrain to Adapt, RoboMem
-
Language-Critique Imitation Learning from Suboptimal Demonstrations Learns from imperfect demos through language critiques, not a single score. Chih-Han Yang, Dai-Jie Wu*, Yun-Ping Huang*, Ping-Chun Hsieh, Kenneth Marino, Shao-Hua Sun CoRL 2026 Workshops · Oopsie, LfCICLR 2027 under review
-
ASAP: Arbitrary-Step Safe Planning with MeanFlow and Control Barrier Functions Safe generative planning in as few as one step, even under unseen constraints. Tzu-Yuan Huang, Yi Yue, Dai-Jie Wu, Stefan Sosnowski, Shih-Hsin Wang, Kenneth Marino, Armin Lederer, Sandra Hirche ICLR 2027under review
-
Intervention-Robust VLA: Diverse Demonstrations for Robot Deployment with Safety Filters Diverse demos keep a VLA working when a safety filter steps in. Tzu-Yuan Huang, Shao-Hsuan Chang, Dai-Jie Wu, Ralf Römer, Yung-Hui Li, Stefan Sosnowski, Angela P. Schoellig, Kenneth Marino, Sandra Hirche ICRA 2027under review
-
SAD-Flower: Flow Matching for Safe, Admissible, and Dynamically Consistent Planning Flow-matching planning that guarantees safety, action limits and dynamics. Tzu-Yuan Huang, Armin Lederer, Dai-Jie Wu, Xiaobing Dai, Sihua Zhang, Hsiu-Chin Lin, Shao-Hua Sun, Stefan Sosnowski, Sandra Hirche ICML 2026arXiv
-
CooT: Learning to Coordinate In-Context with Coordination Transformers Adapts to unfamiliar partners in context, with no parameter updates. Huai-Chih Wang, Hsiang-Chun Chuang, Hsi-Chun Cheng, Dai-Jie Wu, Shao-Hua Sun ICML 2026 · ICML 2025 MAS WorkshoparXiv
-
Diffusion Imitation from Observation Imitation from state-only demos, with a diffusion model as the discriminator. Bo-Ruei Huang, Chun-Kai Yang, Chun-Mao Lai, Dai-Jie Wu, Shao-Hua Sun NeurIPS 2024arXivCode
-
CRAB: Cross-environment Agent Benchmark for Multimodal Language Model Agents A benchmark for multimodal agents whose tasks span desktop and phone. Tianqi Xu*, Linyao Chen*, Dai-Jie Wu*, Yanjun Chen*, Zecheng Zhang, Xiang Yao, Zhiqiang Xie, Yongchao Chen, Shilong Liu, Bochen Qian, Anjie Yang, Zhaoxuan Jin, Jianbo Deng, Philip Torr, Bernard Ghanem, Guohao Li ACL Findings 2025 · NeurIPS 2024 OWA WorkshoparXivCode
-
StyleDiT: A Unified Framework for Diverse Child and Partner Faces Synthesis with Style Latent Diffusion Transformer Pin-Yen Chiu*, Dai-Jie Wu*, Po-Hsun Chu, Chia-Hsuan Hsu, Hsiang-Chen Chiu, Chih-Yu Wang, Jun-Cheng Chen IEEE FG 2026 · oral
-
Towards Validating Face Editing Ability in Generative Models Dai-Jie Wu*, Pin-Yen Chiu*, Chih-Yu Wang, Jun-Cheng Chen IEEE VCIP 2024 · oral
-
A high-throughput pipeline for phenotyping, object detection and quantification of leaf trichomes Andrea González-Muñoz, Dai-Jie Wu, Ana Belén Perera-Rodríguez, Mohamed Rekik, Silvio Giancola, Brande B. H. Wulff, Catherine Gardener Springer TAG 2025
-
Cross-Modal and Cross-Domain Knowledge Transfer for Label-Free 3D Segmentation Jingyu Zhang*, Huitong Yang*, Dai-Jie Wu*, Jacky Keung, Xuesong Li, Xinge Zhu, Yuexin Ma Springer PRCV 2023
Experience
- SPARK Lab, University of UtahJan 2026 – nowGraduate Research Assistant with Prof. Kenneth Marino · Salt Lake City
- SHENNONGSHI.AIMay – Dec 2025Research Intern · TaipeiBuilt the company's first end-to-end robot learning pipeline for cooking, from data collection to real-robot integration.
- Robot Learning Lab, National Taiwan UniversityMay 2024 – Dec 2025Research Assistant with Prof. Shao-Hua Sun · TaipeiLed the lab's first real-world robotics project on ALOHA, from hardware setup and teleoperation to deployment.
- GenAI Center, KAUSTMay 2023 – May 2024Research Engineer with Prof. Bernard Ghanem and Dr. Guohao Li · Thuwal, Saudi Arabia
- CITI, Academia SinicaOct 2022 – May 2023Research Assistant with Dr. Jun-Cheng Chen · remote
- Computer Vision Team, Wisers AI LabAug 2021 – Jun 2022Machine Learning Engineer · remote
Education
- University of UtahPh.D. in ComputingJan 2026 – now
- National Cheng Kung UniversityB.S. Mechanical Engineering & B.B.A. Statistics (dual degree)2016 – 2021
Projects
Honors & Service
Honors
- 2026Top Reviewer, NeurIPS 2026 (top 8%)
- 2026Gold Reviewer, ICML 2026
- 2025PhD Fellowship, Kahlert School of Computing, University of Utah
- 2021Silver Medal, ICCV 2021 Landmark Retrieval Challenge
- 2nd Place, National Electric Vehicle Design and Innovation Challenge
Reviewing
NeurIPS 2026 · ICML 2026 · ICLR 2025, 2026, 2027 · CoRL 2025, 2026 · ICCV 2025
Teaching & Talks
- Aug 2025Robot learning tutorial, ROS CON Taiwan 2025 · Slides
- Fall 2024Teaching Assistant, Reinforcement Learning, National Taiwan University
Outside the lab
I snowboard, climb and hike, and I like to cook, usually trying a dish I haven't made before.
More about me
- SnowBrighton is my favorite resort in Utah; of everywhere I've ridden, Big Sky in Montana is the best.
- ClimbingBouldering indoors and out; Little Cottonwood Canyon in Utah has plenty of options. I've climbed in gyms in Taiwan, Japan, Korea and the US. Favorites so far: 明德原岩 and 市民攀岩館 in Taiwan, and Bouldering Project and The Front in Salt Lake City.
- MountainsAt NCKU I was in the mountaineering club: route scouting, backcountry trekking, canyoneering and river tracing. Country high points: Yushan, Taiwan 3,952 mMount Kinabalu, Malaysia 4,095 mZugspitze, Germany 2,962 mHighest so far: Thorong La, on the Annapurna Circuit 5,416 m
- WaterI used to swim on a swim team. Saudi Arabia was the first desert I visited; I worked there for a year and learned to windsurf and sail.
- TravelI like backpacking and road trips. I've hiked in Taiwan, Malaysia, Germany, Japan, Nepal, Indonesia, Turkey, Bulgaria and Austria; Kamikochi and the Annapurna Circuit stay with me most. In the US, my favorite national park so far is Denali in Alaska, with a glacier landing. Bulgaria is my favorite country in Europe, and I want to see more of Eastern Europe. South America and Africa are next on the list.
- NextTrad climbing and more crags outdoors, ice climbing, learning to ski, more gyms, resorts and national parks, and one day riding the backcountry.
