Kevin Lu

I am an AI resident at Facebook AI Research, advised by Amy Zhang and Yuandong Tian. My research interests are in sequential decision making, as well as universality in deep learning and AI. I previously graduated with my B.S. in Electrical Engineering and Computer Science from UC Berkeley in 2021, where I did undergraduate research as part of the Robot Learning Lab advised by Igor Mordatch and Pieter Abbeel.

Email / CV / Google Scholar / Twitter / Github

Research

I am broadly interested in artificial intelligence. I am particularly interested in making learning more autonomous, such as enabling agents to learn in reset-free, nonstationary, or unsupervised settings; or by incorporating more sources of supervision, such as offline or multimodal data.

	Decision Transformer: Reinforcement Learning via Sequence Modeling Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch arXiv preprint, 2021 arXiv / website / code press (SyncedReview) / press (The Gradient) / video (Yannic Kilcher) / seminar (Eindhoven RL) We show that a simple transformer trained with a sequence modeling objective can perform competitively with strong specialized offline RL methods.
	Pretrained Transformers as Universal Computation Engines Kevin Lu, Aditya Grover, Pieter Abbeel, Igor Mordatch arXiv preprint, 2021 arXiv / blog / code press (The Batch) / press (VentureBeat) / podcast (TWIML) / video (Yannic Kilcher) We show that a transformer pretrained on natural language can, without finetuning of the self-attention and feedforward layers, match the performance of a transformed fully trained on a downstream non-language modality.
	Reset-Free Lifelong Learning with Skill-Space Planning Kevin Lu, Aditya Grover, Pieter Abbeel, Igor Mordatch International Conference on Learning Representations, 2021 NeurIPS Deep RL Workshop, 2020 (Contributed Talk) ICLR Never-Ending Reinforcement Learning Workshop, 2021 (Invited Paper) arXiv / website / oral / poster / code We show planning over a space of skills is a key component of successful reset-free lifelong learning, avoiding sink states, improving stability, and increasing learning signal.
	Efficient Empowerment Estimation for Unsupervised Stabilization Ruihan Zhao, Kevin Lu, Pieter Abbeel, Stas Tiomkin International Conference on Learning Representations, 2021 arXiv / website / code We design a new unbiased empowerment estimator and show it represents empowerment more faithfully than traditional variational mutual information algorithms.
	Adaptive Online Planning for Continual Lifelong Learning Kevin Lu, Igor Mordatch, Pieter Abbeel NeurIPS Deep RL Workshop, 2019 (Contributed Talk) arXiv / website / oral / poster / code Tackling reset-free learning in dynamically changing worlds by combining model-based planning with model-free learning, utilizing an expensive planner only when necessary.

Teaching

EECS 126: Probability and Random Processes
Head Teaching Assistant: Spring 2021, Fall 2020
Teaching Assistant: Spring 2020, Fall 2019

CS 70: Discrete Math and Probability
Reader: Spring 2019

Website template from Jon Barron