Teach Robots New Skills Easily

As long as things stay the same, warehouse robots accurately pick and place for shipment without problems. However, change the process, and you must reprogram the robot and retrain the system.

MIT researchers developed a training technique that requires a handful of human demonstrations and a mere 10 – 15 minutes. Based on a neural network designed to reconstruct the shapes of 3D objects, with just a few demonstrations, the system can grasp new objects similar to those in the demos.

Based on a new neural network model, a Neural Descriptor Field (NDF) can learn the 3D geometry of a class of items. It computes the geometric representation for a specific item using a 3D point cloud, a set of data points or coordinates in three dimensions. The system can then directly apply it to objects in the real world.

In testing, the method had a success rate of 85 percent on pick-and-place tasks with new objects in new orientations. The previous baseline was only able to achieve a success rate of 45 percent.

So far, their method only works for the particular object category in which the robot has trained.