Tony Z. Zhao

I am a third-year CS PhD student at Stanford, advised by Chelsea Finn. I am also a part-time student researcher at Google Deepmind.

Previously, I interned at Tesla Autopilot and Google X Intrinsic. I received my Bachelor's degree in EECS from Berkeley in 2021, advised by Sergey Levine and Dan Klein.

I want to enable robots to perform complex fine manipulation tasks. I am excited about startups and the future of autonomous robots.

Email  /  Twitter  /  LinkedIn  /  Scholar

[Jan 3, 2024] We released the code for Mobile ALOHA 🏄 Hardware and Learning.
[Mar 27, 2023] We released the code for ALOHA 🏖️ and ACT.
[Sep 6, 2022] I am the recipient of Stanford Robotics Fellowship 2022-23.
[Feb 22, 2021] We released the code for contextual calibration: Github link.
[Oct 29, 2020] We released the code for MELD: Github link.

Introduces ALOHA Unleashed 🌋: a simple recipe for robot dexterity. We show that large scale data collection combined with expressive models can be effective in learning challenging bimanual manipulation tasks.

Introduces Surgical Robot Transformer 🪡: learning fine-grained surgical manipulation tasks with imitation learning. Learn tasks such as knot tying, needle pickup, and tissue lift.

Introduces ALOHA 2 🤙: an enhanced version of ALOHA 🏖️ that has greater performance, ergonomics, and robustness compared to the original design.

Introduces Mobile ALOHA 🏄: A low-cost open-source mobile manipulator. Learn to saute a shrimp and use cabinets with 50 demos and co-training with the ALOHA 🏖️ data.

Introduces ALOHA 🏖️: A Low-cost Open-source HArdware for bimanual teleoperation, and ACT: Action Chunking with Transformers. Learn tasks such as opening a translucent condiment cup and slotting a battery with 80-90% success, with 50 demos.

Introduces the Open X-Embodiment Dataset, a dataset from 22 different robots collected through a collaboration between 21 institutions, as well as RT-X models.

A large-scale empirical study on pretrained visual representations, focusing on the distribution shift between pretraining videos and downstream control tasks.

Combines offline meta-RL with online finetuning for industrial insertion. Our method solves 12 new tasks including RAM and network card insertion, with 100% success rate and an average of 6 minutes online interactions.

Introduces contextual calibration, a data-free procedure that improves GPT-2/GPT-3’s accuracy (up to 30% absolute) and reduces variance across different prompt designs.

Towards autonomous robot training. Learning complex dexterous manipulation skills with 16-DOF robotic hand and 6-DOF Sawyer arm, through 60 hours of non-interrupted training.

Demonstrates that predictions of deep NLP models can be manipulated with concealed changes to the training data. Experimented with widely used models (e.g. BERT, GPT-2) and tasks including text classification, language modeling and machine translation.

Bridges Meta-RL for fast skill acquisition and latent state models for state estimation. First meta-RL algorithm trained on real-world robotic control setting from images.

website template