NVIDIA has developed a new deep learning-based system that enables robots to learn tasks by observing human workers. This not only makes it easier to teach robots new tasks, but also speeds up this learning process. During a demonstration of the system, a robot managed to successfully perform a task after just one demonstration.
Traditional industrial robots are programmed for specific tasks, which is a time-consuming process. Industrial robots are capable of performing tasks very efficiently and quickly, but offer little flexibility and cannot be easily redeployed for another task. Also, industrial robots must be placed behind a safety fence to avoid accidents with human workers.
Collaborative robots or cobots are also available, which, thanks to sensors, are able to detect the presence of humans and adjust their behaviour accordingly. As a result, these robots do not need to be placed behind a safety fence and can work more closely with human workers. Unlike traditional robots, collaborative robots often use artificial intelligence and neural networks to learn certain tasks. Tasks can therefore be learned faster, allowing robots to be used more flexibly.
Training collaborative robots
However, training collaborative robots requires a lot of training data, which must be accurately labelled. Creating this data is time-consuming and therefore a major bottleneck in such systems, according to NVIDIA. "For robots to perform meaningful tasks in practice, it must be easy to communicate the task to the robot; this includes the desired outcome and hints that help achieve that outcome," NVIDIA researchers report in a research paper. "With demonstrations, a user can communicate a task to the robot and provide hints on how best to perform the task."
The researchers have addressed this problem by using synthetic data to train neural networks. This data can be generated automatically, making it possible to produce an almost infinite amount of labelled training data in a short time and with minimal investment. The researchers deployed this synthetic data to use NVIDIA TITAN X graphics processors (GPUs) to train a series of neural networks to perform tasks related to interpreting images, creating programmes describing the steps required for specific tasks and executing these programmes. This training allows robots to self-determine how to perform a task based on a demonstrated task. For example, during a field test by NVIDIA, a robot managed to successfully perform a task after just one demonstration.
In addition, the researchers have adopted an "image-based domain randomisation approach". This involves generating highly varied synthetic data that is used to train a robot. The robot's perception network is then fooled so that it interprets real-world data as simply a new variation on its training data.
Detect and fix interpretation problems
Whether a robot manages to successfully perform a demonstrated task depends largely on the robots' observations and how the robot interprets these observations. If errors occur in this, it can affect how the task is performed. To make it easier for users to detect these kinds of errors, the robot, with the help of NVIDIA's technology, creates a human-readable description of the steps it wants to go through to perform the task. This description allows users to quickly detect and correct errors in the robot's interpretation, even before the robot tries to perform the task in practice.
The video below shows the robot's test. This involves a human stacking a number of coloured blocks on top of each other, after which the robot performs the task itself.
NVIDIA's team plans to further explore the use of synthetic training data to train robots in the near future to expand the capabilities of their new training method. More information about the study can be found in a researcher paper released by the researchers.
Author: Wouter Hoeffnagel
Source: NVIDIA
Source: Research paper 'Synthetically Trained Neural Networks for Learning Human-Readable Plans from Real-World Demonstrations'
