Core Machine Learning
- 11/2016 Categorical Reparameterization with Gumbel-Softmax: A method for backpropagation through discrete categorical samples in neural networks.
- 10/2018 WAIC, but Why? Generative Ensembles for Robust Anomaly Detection: We use generative models to perform out-of-distribution detection. Concurrently with Nalisnick ‘18, we show that likelihood models can suffer OoD problems when performing OoD tasks and improve their robustness with uncertainty estimation via ensembling.
- 7/2020 Meta-Learning Requires Meta-Augmentation Data augmentation techniques are ubiquitous in ML and often result in a substantial improvement in performance. We formalize “meta-augmentation” and show that you can apply it to pretty much any meta-learning problem and any meta-learner.
Control and Robotics
- 9/2017 Time Contrastive Networks: We came up with a really data-efficient, unsupervised feature representation for robotic imitation learning. Using this as a state representation, we can get a robot to imitate a human pouring liquids with a single video demonstration.
- 9/2017 Deep Reinforcement Learning for Vision-Based Robotic Grasping: We came up with a simulated benchmark for robotic bin picking from RGB images and evaluated 6 off-policy learning algorithms on it.
- 10/2018 Grasp2Vec: Learning Object Representations from Self-Supervised Grasping: Robots that teach themselves instance grasping without any object labeling.
- 6/2018 QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation: We use reinforcement learning to learn grasping from RGB images on real robots. Robot learns emergent manipulation behaviors like retrial and object singulation!.
- 5/2019 Watch, Try, Learn: Meta-Learning from Demonstrations and Rewards: We train an agent to imitate visual manipulation skills from raw pixels. The agent meta-learns how to re-try from failures, allowing it to fix its own mistakes in a single trial.
- 11/2021 BC-Z: Zero-Shot Task Generalization with Robotic Imitation Learning: We train a mobile manipulation policy to empty bins, open doors, and follow unseen language commands using the simplest, task-agnostic method possible: behavior cloning teleoperated demonstrations from raw pixel observations.
- 3/2022 Practical Imitation Learning in the Real World via Task Consistency Loss: We use the BC-Z network architecture to control a mobile manipulator to open doors. As end-to-end robotic learning scales up to more general settings, real-world evaluation becomes hard. The sim2real gap must be kept small if we are to trust simulated evaluation. This work proposes a novel CycleGAN-based data augmentation regularizer to keep the sim2real gap small.
- 3/2022 Bayesian Imitation Learning for End-to-End Mobile Manipulation: The TCL work above used a self-supervised learning + data augmentation approach for closing the sim2real gap. We investigate a model-centric regularization approach via VIB, which brings up our latched door opening capability to 96% success across 10 different doors.
- 4/2022 Do As I Can, Not As I Say: Grounding Language in Robotic Affordances: Language Models can not only tell jokes, they can also do “long term thinking” for robots! The benefits go both ways; the robot’s value functions ground language sampling to what is realistic in its current context. We scale this to hundreds of mobile manipulation tasks in a real kitchen.
- 5/2022 Multi-Game Decision Transformers: We show that a Decision Transformer (plus modifications) is a better “data sponge” than some online and offline RL algorithms when applied to the difficult task of playing many Atari games at once.
A full list of publications can be found on Google Scholar.