Robota-Embed: Unified VLA Embeddings for Robot In-Context Learning and Evolution
Preprint · 2026
Unified vision-language-action embeddings for retrieving useful robot experience.
University of California, Berkeley
Zekai Wang
I am an undergraduate at UC Berkeley studying Computer Science and Applied Mathematics, with graduation expected in December 2026. I conduct research in Professor Trevor Darrell’s group at Berkeley Artificial Intelligence Research (BAIR).
My research spans robot learning and dexterous manipulation, including tactile feedback, learning from human video, and in-context learning. I also build the data, training, and inference systems that make learning-based policies work on real robots.
Research interests
I am interested in learning from diverse sources of data and building efficient systems for training and deploying foundation models. My work has explored imitation and reinforcement learning, visuo-tactile policies, and the use of human demonstrations and generated video for robot control. I am particularly interested in how model architecture, data pipelines, and execution systems jointly shape a model’s capabilities.
Publications & preprints
* Equal contribution. Longer author lists are abbreviated; linked papers contain the complete lists.
DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation
Preprint · 2026
Extending pretrained video world models to predict visual and multi-finger contact dynamics for dexterous control.
T-Rex: Tactile-Reactive Dexterous Manipulation
CoRL 2026
Combining high-frequency tactile feedback with vision-language-action models for reactive dexterous manipulation.
Contrastive Action-Image Pre-training for Visuomotor Control
Preprint · 2026
Learning action-centric visual representations from egocentric human video.
From Generated Human Videos to Physically Plausible Robot Trajectories
Preprint · 2025
GenMimic translates generated human videos into humanoid motion through pose recovery, retargeting, and physics-aware reinforcement learning.
In-Context Learning Enables Robot Action Prediction in LLMs
ICRA 2025
RoboPrompt enables off-the-shelf text-only LLMs to predict robot actions from in-context demonstrations without additional training.
Selected systems & projects
Real-time dexterous robot systems
I designed a bimanual teleoperation system with tactile feedback and built multimodal data collection and processing pipelines. My inference work includes real-time action chunking and CUDA graph capture for smooth, reactive policy execution.
T-Rex project →Asynchronous reinforcement learning
I implemented an online multi-GPU asynchronous RL training system, using dedicated KV-cache producer GPUs to reuse invariant model prefixes and reduce update latency.
Ongoing research · see CV for backgroundCourse projects
Computer vision and graphics coursework.
Teaching & experience
I served on Berkeley course staff for five semesters, including as Head TA for DATA 100. I was also a Software Engineer Intern at Google in Summer 2025, working on LLM agent evaluation and tracing tools.
- DATA 182 · Introduction to Deep Learning — TA, Spring 2026
- EECS 126 · Probability and Random Processes — TA, Spring 2025
- CS 189 · Introduction to Machine Learning — Tutor, Fall 2024; TA, Spring 2024
- DATA 100 · Principles and Techniques of Data Science — Head TA, Summer 2024