How Machine Learning Is Developing Robotics

How Machine Learning Is Developing Robotics

Picture of Factored Data Experts
Factored Data Experts

Where are we at with robotics? What are the biggest challenges robotics is facing today?

In this post, we’ll first take a look at the robotics industry’s latest trends and challenges. We’ll also look at how Machine Learning (ML) can be applied in the real world in conjunction with robotics. Last but not least, we’ll provide a recent application of ML in the field of robotics that our Factored engineers did.

Let’s get to it!

The Robotics Industry: Where We’re At

As the Robotics Business Review recently noted, combining Machine Learning and 5G connectivity is a natural next step in the development of more autonomous robots. However, this requires the availability of robust robotics development and execution platforms—as well as robotics engineering expertise.

Therefore, even though more and more technology providers are offering robotics development ‘platforms’ that incorporate ML and 5G enablers, “development platforms alone are not enough to drive robotics innovation forward—at least not until the supply of robotics engineering talent exceeds demand.”

In fact, the significant shortage of engineers in robotics and Artificial Intelligence is hampering more sophisticated development and research.

Applying Machine Learning to Robotics

Machine Learning is a promising source of solutions to many robotics and industrial control problems. Supervised Learning and Reinforcement Learning can help control both the actions of robotic elements and the important variables of industrial processes. This is by producing dynamical models coupled with classical control techniques or by directly producing data driven controllers.

Much of the current Machine Learning exploration in system control is directed towards regulation. That is, making a system reach a state with little regard to the path it followed to get there. But in many applications—and especially in robotics—the trajectory followed by the system is just as important as the final destination. This requires a more sophisticated type of control.

Think of a robotic arm that is drawing or painting something; the arm has to accurately follow a very specific trajectory for the final result to meet the requirements. Classic Control Theory has studied trajectory control for a long time, and it usually assumes there’s a good dynamical model of the system to use to design controllers.

However, it’s not always easy to get a dynamical model of a complex system just from mathematics and first principles. This is where Machine Learning plays a huge role. That’s because Machine Learning can approximate a good model or even produce a controller without the need for a dynamical model at all.

Putting It In Practice

Recently, a team of Factored engineers participated in the robotic track of the Learning by Doing competition of NeurIPS 2021. This track of the competition asked participants to design controllers for different kinds of simulated robotic arms so that their effectors (robotic arms’ ends) follow a desired trajectory.

The team of Factored engineers used simple techniques based in Linear Algebra, Statistics, and Machine Learning to propose controllers for these robotic arms. The team achieved third place in the competition. They used concepts and techniques like linear regression, polynomial regression, covariance, and singular value decomposition.

One of their findings was that properly handling unrealistic predictions for the control actions was very important. They used statistics and linear algebra to implement a smarter clipping strategy when the predicted control action was far away from observed data.

Admittedly, their results were mixed. They were able to follow some trajectories really well, while others threw off their controllers. The fact that they reached third place while failing in some of the trajectories is telling of how difficult the task was. In general, they found that data-driven control of robotic elements is still challenging when the task is trajectory control as opposed to simple regulation.

The Bottom Line

Machine learning and robotics are set to change multiple industries—but only if they have the talent they need to put these technologies in place. Thankfully, Factored has the expertise to help companies be at the forefront of this change.

Book a meeting today to discuss how Factored machine learning engineers can help you augment your processes with machine learning.

Related Posts