Action Chunking (ACT)

Action Chunking with Transformers (ACT) is an imitation learning algorithm that enables low-cost robots to perform precise, fine-grained manipulation tasks by predicting sequences of future actions rather than single steps.

Developed by researchers at Stanford and Google, Action Chunking with Transformers (ACT) solves the compounding error problem in robotic imitation learning (1.1.4, 1.2.9). Instead of predicting one action at a time, ACT uses a Conditional Variational Autoencoder (CVAE) and a Transformer architecture to output a continuous chunk of future actions (1.1.4, 1.1.9). This approach allows low-cost, imprecise hardware to master highly complex, contact-rich tasks (such as opening condiment cups, slotting batteries, or folding clothes) with 80% to 90% success rates using only 10 minutes of human demonstration data (1.2.2, 1.2.4, 1.2.9).

https://tonyzhaozh.github.io/aloha/