2025

1. 

   Imitation Learning from a Single Temporally Misaligned Video

   William Huey^*, Huaxiaoyue Wang^*, Anne Wu, Yoav Artzi, and Sanjiban Choudhury

   In International Conference on Machine Learning (ICML), 2025

   TLDR arXiv Code

   To learn from a single video, current approaches leverage an image encoder, distance metric on the encoder’s latent space, and distribution-matching algorithm to produce dense rewards for RL. We prove that frame-level distribution matching algorithms cannot accurately model the ground truth reward of a video demonstration. We propose ORCA (ORdered Coverage Alignment), an intuitive algorithm that matches distributions at the sequence level, achieving better performance on 16 video-imitation tasks in two simulation environments.

2024

1. 

   Time Your Rewards: Learning Temporally Consistent Rewards from a Single Video Demonstration

   Huaxiaoyue Wang^*, William Huey^*, Anne Wu, Yoav Artzi, and Sanjiban Choudhury

   In Conference on Robot Learning (CoRL) Workshop on Whole-body Control and Bimanual Manipulation, 2024

   TLDR OpenReview Code

   We show that existing approaches for inferring reward functions from video demonstrations result in temporal inconsistencies: agents complete subtasks in the wrong order, get stuck along the expert trajectory, or fail to stay in the final state. Our approach mitigates this problem, solving humanoid control tasks quickly and efficiently.

2024

1. 

   Distilling Vision-Language Models for Real-Time Traversability Prediction

   William Huey^*, Sean Brynjolfsson^*, and Donald Greenberg

   Cornell Discover Undergraduate Research in Engineering Showcase, 2024

   TLDR PDF Code Poster

   We show that large pretrained vision language models can accurately predict traversability. We distill fast traversability prediction models that run in real time on robot hardware, allowing for long horizon unguided exploration.

2023

1. 

   SuperEdge: A CNN-Based Approach for Quadruped Traversability Analysis from Incomplete Observations

   William Huey

   CS 4756: Robot Learning Final Project, 2023

   TLDR PDF Code Video

   I demonstrate self-supervised traversability learning in simulation given sparse point cloud observations. To support this project, I developed a navigation stack for legged robots in Nvidia Isaac Sim, which includes traversability estimation, environment graph generation, graph search, and path tracking.

2022

1. 

   Practices in Combinatorial Testing Relevant to Spacecraft Software Verification

   William Huey, D. Richard Kuhn, and William Stanton

   NASA Summer Intern Symposium, 2022