Technology

Transformers (ACT)

Action Chunking with Transformers (ACT) is an imitation learning algorithm that allows low-cost robots to master precise, bimanual manipulation tasks from just minutes of human demonstration.

Developed by researchers at Stanford, Action Chunking with Transformers (ACT) solves the compounding error problem in robotic imitation learning by predicting sequences of future target joint positions (typically 100 timesteps at 50 Hz) rather than single, immediate actions. The lightweight, 80-million-parameter architecture combines a ResNet-18 vision backbone, a Conditional Variational Autoencoder (CVAE) to handle human demonstration variability, and a DETR-inspired transformer decoder. By grouping actions into cohesive chunks, ACT enables low-precision hardware to achieve 80% to 90% success rates on complex, contact-rich tasks (such as slotting batteries or opening translucent condiment cups) with as little as 10 minutes of demonstration data.