Teaching Bots Learn by Watching Human Behavior - Ep. 67

Error Correction

In the realm of robotics, error correction is a crucial aspect of developing autonomous systems. Animesh Garg explains how robots can learn from their mistakes, allowing them to recover and reteach themselves during tasks. For instance, if a robot spills water while cooking, it can go back and boil the water again, showcasing its ability to rectify errors 1. This capability is part of neural task programming, which enables robots to observe human actions and create plans to execute tasks, even correcting their own mistakes along the way 2.

The system can go back and fix those. Or at least that's the idea.

--- Animesh Garg

   

Neural Programming

Neural task programming is a groundbreaking approach that combines modularity and deep learning to enable robots to learn tasks from human demonstrations. Animesh Garg highlights how this method allows robots to convert video demonstrations into sequences of actions, facilitating task execution in real-world scenarios 2. Marynel Vázquez discusses the Jackrabot project, which focuses on social robot navigation, teaching robots to move through crowded spaces without disrupting people 3.

The novelty, or really what is the radical here, is being able to take in a video and convert that into a sequence of primitive actions.

--- Animesh Garg