One ACT Play: Single Demonstration Behavior Cloning with Action Chunking Transformers

TL;DR

This paper introduces a method for one-shot behavior cloning in robotics by augmenting a single demonstration with linear transforms and improving action chunking robustness through a novel temporal ensembling enhancement.

Contribution

It presents a new approach combining demonstration augmentation and an improved action chunking method for effective one-shot learning in robot manipulation tasks.

Findings

Successful learning of three block manipulation tasks from a single demonstration

Enhanced robustness and performance through the novel temporal ensembling method

Demonstrated ability to generalize across varied initial conditions

Abstract

Learning from human demonstrations (behavior cloning) is a cornerstone of robot learning. However, most behavior cloning algorithms require a large number of demonstrations to learn a task, especially for general tasks that have a large variety of initial conditions. Humans, however, can learn to complete tasks, even complex ones, after only seeing one or two demonstrations. Our work seeks to emulate this ability, using behavior cloning to learn a task given only a single human demonstration. We achieve this goal by using linear transforms to augment the single demonstration, generating a set of trajectories for a wide range of initial conditions. With these demonstrations, we are able to train a behavior cloning agent to successfully complete three block manipulation tasks. Additionally, we developed a novel addition to the temporal ensembling method used by action chunking agents during inference. By incorporating the standard deviation of the action predictions into the ensembling method, our approach is more robust to unforeseen changes in the environment, resulting in significant performance improvements.