Machine Learning, Deep Learning & Neural Networks: Part 3

Machine Learning is being hailed as a game changer for robotics & automation for many reasons but in particular for the Vision, Picking, Movement and Data areas.

Neural Networks in Automation 

Applications already making an impact

Machine Learning is very useful for vision technology – where camera guided robot automation is required. Machine Learning, when combined with automation is useful in areas such as Picking where a robot may come across new objects regularly or where programming a robot for all possible scenarios is impractical. Materials Handling and Industrial Automation will both see huge benefits from incorporating Machine Learning into systems, helping streamline and improve existing practices within those industries. Data is the last major area that will see huge impact from Machine Learning, as it will streamline automated processes in understanding and processing huge amounts of data, as well as finding patterns and trends.


Bin Picking RGB Data & Detection

Bin Picking Robot from Scott, uses RGB data & Machine Learning to detect and pick brackets


Being able to make decisions in real time is one of the big reasons why Machine Learning will have such a huge influence on how automation is approached in a variety of industries. In healthcare, symptoms and patient data can be input for analysis and then a doctor can review the findings, speeding up complex and delicate diagnostics. Neural Networks have also been shown to perform better in detecting a range of illnesses in scans when trained correctly. Care industries will also see advanced assistive technology such as companion robots. Machine Learning means these robots will be able to learn and respond better to their environments providing a higher level of assistance and care for those they help.

The ability to adapt to make appropriate & accurate decisions based on diverse data and inputs is also what makes Machine Learning such a valuable asset in Materials Handling. Complex production lines and other variable processes can be automated more efficiently with less error margins and down time. For example, a robot working in a cold storage room picking & placing boxes may come across boxes with varying degrees of frost or ice. Machine Learning models trained on a large data set will be able to cope with the diversities presented which enables the robot to recognize the unfamiliar shape as product and then correctly pick the box.


Vision system is used for data collection in forequarter cut paddywack detection

Vision system is used for data collection in forequarter cut paddywack detection


Scott has used Machine Learning in large-scale meat processing. In this area, Neural Networks outperformed classical vision techniques because the models were capable of adapting to the immense diversity of data in the meat industry, improving outcomes for producers. One such use was in predicting the position of a specific cut on a piece of meat. This cut was more difficult than most as its position is often partially obscured from the field of view of the vision system because of the diversity in the size and shape of the meat. These challenges meant that creating this cut accurately, especially in an automated environment, was difficult. However, by using a Neural Network trained on a diverse data set, the system has a cutting accuracy never before seen in the industry.


Rib Prediction with Convolutional Neural Network

Convolutional Neural Network predicts rib placement using X-Ray data for accurate cutting


Similar Machine Learning was used in combination with vision technology for a mining application in which Scott was involved. This application saw Neural Networks trained to distinguish between contaminants and gold ore. The machine learnt the difference between ore pixels and other pixels, helping it determine through analysis of the data, which pixels were statistically more likely to be ore and those more likely to be contaminants. Machine Learning is ideal for this application because of both the massive diversity and the amount of data available for training. Once trained it could correctly detect in real time, deciding what was and was not contaminant, something that would be impractical using traditional Computer Vision techniques. The resulting decisions and data of this application could additionally work as future training, reinforcement for further detections.


Contaminant Detection in Gold Ore

Contaminant detection in gold ore, contaminant's shown with a red outline


While we have yet to fully realise or unleash the full potential of Machine Learning in industrial robotics and automation, advances in detection, analysis and problem solving have already simplified processes in a multitude of industries including materials handling, meat processing and mining. Applications that utilise Machine Learning currently are already making an impact and the development and learnings gained from these existing systems will shape how we use Machine Learning/Deep Learning and Neural Networks in robotics & automation going forward - making the future both promising and exciting.


Missed out on our Machine Learning series? Catch up with part 1 Neural Networks, Machine Learning and how they work and part 2 How to Train a Network & Network Types


For more resources on Neural Networks in Automation see:

Machine Learning in Robotics – 5 Modern Applications by EMERJ, an AI Research & Advisory Company

Applying Artificial Intelligence and Machine Learning in Robotics from the Robotics Industries Association

Entering a New Robotics Age with Machine Learning by Electronic Design


Mailing List