When getting started with machine learning in robotics, it can be difficult to select the best technique for your use case. The ML field overflows with diverse techniques, each of which excels when applied to the right type of problems.
Here are the five primary ML techniques used in the robotics field—and how to decide on the best approach for your challenge.
Supervised learning is the first and most popular technique that most people consider when they think about ML. In supervised learning, algorithms are fed labeled data in the form of input and output pairs. The output labels inform the model of the outcomes it’s intended to predict.
Supervised learning is a popular topic of research in the ML community as a whole, with recent innovations often exploring the development of more advanced deep neural networks. Because deep learning networks perform best when given large quantities of training data, these models have become increasingly appealing in this data-dense world. When high-quality labeled data is available for training, supervised learning can provide powerful insights across a variety of fields.
In robotics, this technique is most useful for applications that include well-labeled input data and a clearly defined targeted outcome. For example, supervised learning is frequently the technology behind object detection. In manufacturing, industrial robots can be trained to identify specific mechanical parts, which can be useful for detecting item locations or counting available objects. However, because robots typically produce large amounts of complex, unlabeled data, supervised learning is generally less popular in robotics than in other ML-related fields.
Unsupervised learning includes algorithms that learn from unlabeled data. As unsupervised learning techniques have become more advanced, they’ve unlocked opportunities to learn from data that is difficult to label. In robotics, where robots and their sensors are exposed to immense amounts of unlabeled data, these techniques provide a way to extract insights from data that was previously considered extraneous.
Unsupervised learning can help uncover new ways to represent and structure data, making it a suitable technique for analyzing data for previously unidentified patterns. In robotics, unsupervised learning is especially useful for anomaly detection, since the technique makes it easy to detect parts that don’t quite match with the others. This makes it a great fit for identifying defective parts.
Because unsupervised learning is useful for detecting patterns, it can also be used to prepare data for supervised learning by generating label clusters or other identifiers.
With reinforcement learning, a model takes actions to maximize cumulative reward. While this technique has a preferred outcome, it does not use labeled data and does not require suboptimal actions to be corrected. Instead, reinforcement learning allows the model to autonomously discover an optimal behavior through trial and error.
This process involves balancing the exploration of unseen options with the knowledge of the options that have already been pursued to seek an optimal outcome. The model is provided with the rules for the reward mechanism and aims to achieve an outcome that maximizes that reward.
Reinforcement learning is used heavily in the field of robotics as the primary technique behind path planning. In path planning, a robot’s goal is to find the shortest, most obstacle-free path from a start point to an end point. Reinforcement learning allows the robot to freely explore the space to determine an outcome.
While this is a well-developed field in robotics, it’s still a popular topic for research, with scientists exploring new and improved techniques for information representation and path exploration.
Self-supervised learning falls between supervised and unsupervised learning. With this technique, models learn from unlabeled data with the goal of generating classified outputs. This makes it possible to perform a sort of supervised learning without actually having labeled data.
This technique is incredibly popular in robotics because it leverages the vast amounts of available, unlabeled data to create labeled outcomes. When working with low-quality datasets, this technique can be useful for data creation and augmentation.
There are, however, some limitations. For self-supervised learning to be a viable technique, the structure of the problem needs to be well known and consistent so that a standard labeling mechanism can be implemented.
In many well-defined use cases, however, self-supervised learning has taken the lead role in robotics. For example, this technique is useful for manufacturing tasks such as pick-and-place, where a robot must pick up a specific object and move it to the appropriate location.
While self-supervised learning is applicable only for well-defined use cases, it’s a popular topic for research in the robotics community. It may provide ways to improve on existing approaches to things such as object manipulation, where reinforcement learning is currently standard practice.
With imitation learning, the model learns by imitating a trainer. The trainer demonstrates the correct action, and when the model repeats the gesture, it receives feedback based on its performance. Because the trainer must perform tasks in a way that the model can replicate, imitation learning works best when the trainer can easily and naturally demonstrate actions to the model. For example, to train a robot with limited degrees of freedom, the trainer should be able to demonstrate motions that align with those degrees of freedom.
There tends to be a great deal of overlap between the types of problems that can be solved via imitation learning and reinforcement learning. Training a model using reinforcement learning, however, typically requires far fewer examples than training a model with reinforcement learning, because imitation learning uses high-quality, dense data that demonstrates an ideal performance.
In robotics, imitation learning is popular for training tasks that mirror human behaviors and that require a lot of precision. For example, imitation learning provides a way to train a robot to grasp objects with a high degree of accuracy and the appropriate amount of force. Picking up new objects is a notoriously difficult task in robotics, so some researchers have postulated that imitation learning may be the future of robotic manipulation.
Because imitation learning requires physical demonstrations, however, it can be a time-consuming ML approach. Developing more efficient frameworks for introducing data to robotic systems may make imitation learning a more viable option beyond research purposes in the future, improving the ability of robots to achieve complex object manipulations.
Many ML challenges in robotics can pursue multiple viable techniques. More complex use cases often use a combination of different learning techniques. When setting up an ML system in robotics, however, there are several factors that you can focus on when selecting your default approach.
When working with well-labeled data and a clear outcome, supervised learning tends to be the ideal approach, while unsupervised learning excels at detecting patterns in unlabeled data. Self-supervised learning, which produces classified outcomes based on unlabeled data, falls somewhere in the middle. This approach can be applied only to well-defined problems, but it’s rapidly becoming a popular technique within the robotics community.
Reinforcement learning and imitation learning, on the other hand, are best for cases where the model must learn the best path to an outcome. With reinforcement learning, the task is expressed in terms of cost and reward, and the model learns through exploration. With imitation learning, the task is displayed by a trainer, and the model learns through repetition and feedback. But you need fewer examples to train a model using reinforcement learning.
Traditionally, reinforcement learning has been the backbone of ML in robotics, especially when it comes to common tasks such as motion and path planning. However, current trends suggest a move toward self-supervised approaches as the mechanisms behind common robotics challenges become more well defined. Imitation learning also shows promise for the development of more precise motion controls. As more complex ML techniques are developed, the possibilities of robotics learning will only grow.