Photo courtesy of HP
“Machine learning and robotics are a perfect match,” suggests HP Fellow Will Allen.
Although experts in the one field rarely stray into the other, Allen says, their potential synergies are real. “Machine learning is very applicable to robotics, and robotics—by which I mean working with physical robots—needs some of the things that machine learning is good at,” he argues.
Now Allen, who has a background as a distinguished innovator in imaging and printing technologies, is co-leading a research team with colleague David Murphy in HP’s Emerging Compute Lab that aims to understand, and potentially harness, those synergies to create a new generation of what the team are calling “Smart Machines.”
One of the main challenges in robotics—where you want electro-mechanical machines to perform specific tasks with some degree of autonomy—is to have the machines move both precisely and efficiently in 3D space. Robots can be programed to accomplish this by mapping a space and then detailing how the robot should move through it—to enter a room with a pile of paper cups, for example, and then place the cups on a counter. But that plotting process is both laborious and requires expertise, and its results can’t easily be applied to very similar tasks that happen to be located in a different space, or even in the same space when the environment is not static.
Several Smart Machines projects, however, are demonstrating that machine learning can both speed up and simplify the programming process, helping make it applicable by non-experts in multiple use cases.
One Labs project is helping improve the inspection process for HP’s large format printers. As soon as each printer is built, the manufacturing group takes a set of nearly 200 reference pictures that record the machine’s exact physical state prior to shipping. This supports both consistency and troubleshooting should an issue occur with a specific printer or set of printers once they are installed.
HP’s Large Format Printer unit wanted to give their inspectors more time to check each printer by automating the image capture process, and they turned to the HP Labs team to help. A conventional approach would plot an inspector’s exact position as he or she took each of the pictures. But the Labs researchers are instead taking a CAD model of the printer’s exterior and an example of every reference picture and then running a machine learning algorithm to figure out how to cause the robot to take those exact photos from those exact same vantage points.
This employs a machine learning technique known as reinforcement learning: A software agent takes an action in an environment, here changing the position of a robot’s camera and taking a picture, and the result is evaluated based on the outcome of that action. When the camera’s positioning captures an image similar to a desired reference image, a positive “reward” is assigned. Over a large number of training cycles, the reinforcement learning algorithm becomes efficient at directing the robot to achieve a positively rewarded outcome.
“You reach a solution by defining what a successful outcome is, not by defining how the robot actually solves the problem,” notes team co-leader David Murphy, who comes to the project from a machine learning background.
Running large numbers of trials with robots in a physical environment can be slow and expensive, Murphy adds. So the algorithm here is trained in a simulator, where, over a number of trials, a model of the robot learns to take pictures from the precise angle required. When fully trained, the final set of directions, known as a policy, will be exported to a physical robot set up to photograph a real printer.
Crucially, this process will deliver a robotics solution that the HP printer team can apply to inspecting other printer models without ever needing to be robotics experts themselves.
When you don’t need deep expertise to either define a robotics challenge or create the best solution to it, new uses for robots immediately open up. “You can now be an office assistant, or brain surgeon, or a hotel manager, or an electrical engineer and still have a robot solve a problem for you,” Allen says.