Albert Yu

Hi there, I'm Albert, a final year PhD candidate at UT Austin working on interesting ways to enable natural language processing to improve robot learning and robot capabilities. I am co-advised by Profs. Roberto Martín-Martín and Ray Mooney.

I did my undergrad at UC Berkeley where I did research with Prof. Sergey Levine on real-world offline RL.

Email: ude.saxetu@uytrebla

Publications

Mixed-Initiative Dialog for Human-Robot Collaborative Manipulation

ICRA 2026

Oral Presentation, Top 4 of 54 accepted papers, 2026 Texas regional NLP Symposium

web pdf poster video X

@inproceedings{yu2026micobot,
  title = {Mixed-Initiative Dialog for Human-Robot Collaborative Manipulation},
  author = {Yu, Albert and Li, Chengshu and Macesanu, Luca and Balaji, Arnav and Ray, Ruchira and Mooney, Raymond and Mart{\'\i}n-Mart{\'\i}n, Roberto},
  booktitle = {International Conference on Robotics and Automation (ICRA)},
  year = {2026},
}

Robots can’t do everything. How can human-robot teams leverage bidirectionally initiated dialog to communicate and collaborate better on long-horizon mobile manipulation tasks?

BUMBLE: Unifying Reasoning and Acting with Vision-Language Models for Building-wide Mobile Manipulation

Rutav Shah | Albert Yu | Yifeng Zhu | Yuke Zhu | Roberto Martín-Martín

ICRA 2025

web pdf code video X

@article{shah2024bumble,
  title = {BUMBLE: Unifying Reasoning and Acting with Vision-Language Models for Building-wide Mobile Manipulation},
  author = {Shah, Rutav and Yu, Albert and Zhu, Yifeng and Zhu, Yuke and Mart{\'\i}n-Mart{\'\i}n, Roberto},
  booktitle = {2025 IEEE International Conference on Robotics and Automation (ICRA)},
  year = {2025},
  organization = {IEEE},
}

How can service robots perform building-wide mobile manipulation? We propose a framework integrating VLM reasoning, long and short-term memory, and a wide spectrum of gross-to-fine motor skills.

Natural Language Can Help Bridge the Sim2Real Gap

Albert Yu | Adeline Foote | Ray Mooney | Roberto Martín-Martín

RSS 2024

Oral Presentation, Top 4 of 28 accepted papers, RSS 2024 Workshop on Semantic Reasoning and Goal Understanding in Robotics

Oral Presentation, ACL 2024 Spatial Language Understanding and RoboNLP Workshop

web pdf poster code video X

@inproceedings{yu2024lang4sim2real,
  title = {Natural Language Can Help Bridge the Sim2Real Gap},
  author = {Yu, Albert and Foote, Adeline and Mooney, Raymond and Mart{\'\i}n-Mart{\'\i}n, Roberto},
  booktitle = {Robotics: Science and Systems (RSS)},
  year = {2024},
}

Language links sim+real visual features via semantic similarity, improving sim2real transfer for visuomotor policies. Language can bridge the sim2real gap even when there are cheaply simulated deformable objects.

Using Both Demonstrations and Language Instructions to Efficiently Learn Robotic Tasks

Albert Yu | Ray Mooney

ICLR 2023

web pdf poster code X

@inproceedings{yu2023deltaco,
  title = {Using Both Demonstrations and Language Instructions to Efficiently Learn Robotic Tasks},
  author = {Yu, Albert and Mooney, Raymond J.},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year = {2023},
}

Teaching a robot simultaneously with both visual demonstrations and language instructions greatly improves sample efficiency and generalization to new tasks.

ARIEL: Leveraging Prior Data to Automate Robotic Reinforcement Learning

CoRL 2022

web pdf X

@inproceedings{walke2022ariel,
  title = {Don't Start From Scratch: Leveraging Prior Data to Automate Robotic Reinforcement Learning},
  author = {Walke, Homer and Yang, Jonathan and Yu, Albert and Kumar, Aviral and Orbik, J{\k{e}}drzej and Singh, Avi and Levine, Sergey},
  booktitle = {Conference on Robot Learning (CoRL)},
  year = {2022},
}

We tackle three challenges for real-world RL—human resets, data efficiency, and generalization—by pretraining on prior data with offline RL, then fine-tuning online with a forward policy that does the task and a backward policy that resets the environment itself.

Parrot: Data-Driven Behavioral Priors for Reinforcement Learning

ICLR 2021

Oral Presentation (Top 1.8% of ICLR submissions)

web pdf X

@inproceedings{singh2021parrot,
  title = {Parrot: Data-Driven Behavioral Priors for Reinforcement Learning},
  author = {Singh, Avi and Liu, Huihan and Zhou, Gaoyue and Yu, Albert and Rhinehart, Nicholas and Levine, Sergey},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year = {2021},
}

How can RL policies explore more efficiently? We learn a flow-based behavioral prior from past successful trials on previous tasks to more efficiently explore on new tasks.

COG: Connecting New Skills to Past Experience with Offline Reinforcement Learning

CoRL 2020

Oral Presentation, 2020 NeurIPS Offline RL Workshop

web pdf code X

@inproceedings{singh2020cog,
  title = {COG: Connecting New Skills to Past Experience with Offline Reinforcement Learning},
  author = {Singh, Avi and Yu, Albert and Yang, Jonathan and Zhang, Jesse and Kumar, Aviral and Levine, Sergey},
  booktitle = {Conference on Robot Learning (CoRL)},
  year = {2020},
}

How can robots generalize to new scenarios outside of their training data? We leverage large prior datasets through offline RL to chain old skills into new ones with only sparse rewards.