Microsoft announced the public preview of Bonsai project, a platform for building autonomous industrial control systems, during the Build 2020 online conference. The company also presented an experimental platform called Project Moab, designed to familiarize engineers and developers with the functions of bonsai.
Project Bonsai is a "machine learning service" that combines machine learning, calibration and optimization to give control systems at the heart of robot arms, bulldozer blades, forklifts, underground drills, rescue vehicles, wind and solar parks and more autonomy. Control systems form a core component of machines in various sectors such as manufacturing, chemical processing, construction, energy and mining, and help with the management of substations and HVAC systems through to fleets of factory robots. The development of AI and machine learning algorithms – algorithms that could address processes that were previously too difficult to automate – requires specialist knowledge.
Project Bonsai tries to combine this expertise with a powerful simulation toolkit hosted on Microsoft Azure.
Ramp up industries
At a high level, Project Bonsai's goal is to accelerate the introduction of Industry 4.0, an industrial transformation that Microsoft defines as the infusion of intelligence, connectivity, and automation into the physical world. Beyond the new technology, Industry 4.0 brings with it new ecosystems and strategies that use AI for a big profit. Microsoft cites a study by the World Economic Forum, which found that 50% of companies that deal with AI within the next seven years could double their cash flow.
For manufacturers in the transition phase, the end goal is often to receive "prescriptive" information, in which adaptive, self-optimizing technologies and processes help devices and machines to adapt to changing inputs and conditions. Existing control systems have a limitation in that they operate with a series of deterministic instructions in predictable, unchangeable environments. Next-generation control systems tap AI to go beyond basic automation, adapt in real time to changing environments or inputs, and even optimize for multiple goals.
The Bonsai project was developed to create these systems, which also use a combination of digital feedback loops and human experience to inform measures and recommendations. Historical data leads to certain processes and product improvements and enables systems to perform tasks such as calibration faster and more precisely than human operators.
Machine lessons and simulation
The Bonsai project is the result of Microsoft's acquisition of Bonsai in Berkeley, California in 2018, previously funded by the company's venture capital area M12. Bonsai is the brainchild of former Microsoft engineers Keen Browne and Mark Hammond, who is now General Manager for Business AI at Microsoft. The couple developed an approach to Google's TensorFlow framework that abstracts low-level AI mechanics and enables subject matter experts to train autonomous systems to achieve goals – regardless of AI suitability.
In September 2017, Bonsai set a new benchmark for autonomous industrial control systems and successfully trained a robotic arm to capture and stack blocks in the simulation. It ran 45 times faster than a comparable approach from Alphabet & # 39; s DeepMind.
Microsoft describes the abstraction process as machine teaching. The central client solves problems by dividing workloads into simpler concepts (or sub-concepts) and then training them individually before combining them. This technique is also called hierarchical deep reinforcement learning when the AI learns by making decisions and receiving rewards for actions that bring it closer to a goal. The company claims that this technology can reduce training time and allow developers to reuse concepts.
For example, in a warehouse and logistics scenario, an engineering team could use machine lessons to train autonomous forklifts. Engineers would start with simpler skills, such as aligning to a pallet, and building on the fact that they would teach the forklift truck to drive onto the pallet, pick it up, and put it down. Ultimately, the autonomous forklift would learn to recognize other people and devices and return to its charging station.
"There is a joke about reinforcement learning among researchers that looks something like this: If you have a problem and model it like a reinforcement learning problem, you now have two problems," Microsoft CVP Gurdeep Pall told VentureBeat in a phone interview. "It's a very complex field. It's not just about choosing the right algorithm – continuous versus discrete, on-policy versus off-policy, model-based versus model-free and hybrid models – it's about rewards."
As Pall explained, rewards in learning reinforcement describe every right step that an AI tries. Creating these rewards – which need to be expressed mathematically – is difficult because they have to capture every nuance of multi-level tasks. And improperly designed rewards can lead to catastrophic forgetting, in which a model completely and abruptly forgets the previously learned information.
“Machine education creates many of these difficult problems and really puts the problem on the rails. This limits the indication of the problem, ”added Pall. “The (Bonsai platform) automatically selects the algorithm and (parameters)… from a whole range of options and has abstraction goals where a user does not have to provide a reward, but has to specify the desired result. Given a state space and this result, we automatically determine a reward function that we use to train the reinforcement learning algorithm. "
Project Bonsai's general-purpose learning platform coordinates the development of AI models. It provides access to algorithms and infrastructure for both model deployment and training, and enables models to be deployed on-premises, on the device, or in the cloud with support for simulators such as MATLAB Simulink, Transys, Gazebo, and AnyLogic. (For local deployments, a controller companion must be connected to the controller computer in real time.) Bonsai customers can use a dashboard to view all active jobs – so-called BRAINs – as well as their training status and options for debugging, checking and refining models. And they can work with colleagues to create and deploy new models together.
It is a largely straightforward process. After concepts with Inkling, Project Bonsai's special programming language, have been programmed into a model, the code is combined with a simulation of a real system and fed into the Bonsai AI Engine for training. The engine automatically selects the best algorithm for training a model, creates the neural networks and adjusts their parameters. The platform runs multiple simulations in parallel to shorten training time and transfer predictions of trained models to software or hardware through libraries provided by Bonsai.
Bonsai takes a “digital twin” approach to simulation – an approach that has become more important in other areas. Based in London, SenSat helps customers in the construction, mining, energy and other industries, for example, to create location models that are relevant to projects they are working on, and translates the real world into a version that is understood by machines can be. GE offers technologies that enable companies to model digital twins of actual machines, the performance of which is closely monitored. Oracle offers services based on virtual representations of objects, devices and working environments. And Microsoft itself offers Azure digital twinsthat models the relationships and interactions between people, places, and devices in simulated environments.
For example, within the Project Bonsai platform, a model learning to control a bulldozer would receive information about the variables in the simulated environment – such as the type of dirt or the proximity of people walking nearby – before it did decides on measures. These decisions would improve over time to maximize the reward, and domain experts could optimize the system to get a working solution.
It's similar to Microsoft's AirSim framework for Unity, which is said to use machine learning to simulate realistic physics environments for testing drones, cars, and more. Like the Project Bonsai platform, it is intended to serve as a safe, repeatable test field for autonomous machines – in other words, as a means of collecting data before prototyping in practice. In a recent technical article, Microsoft researchers showed how AirSim can be used to train and transfer drone-driving AI from simulation to the real world to bridge the gap between simulation and reality.
According to Microsoft, bonsai simulations hosted on Azure can replicate millions of different real-world scenarios that a system may encounter, including edge cases such as sensor and component failures. After training, models can either be used as decision support, integrating them into existing monitoring software to provide recommendations and predictions, or with direct decision-making powers so that the models develop solutions to challenging situations.
For onboard engineers and developers who want to experiment with bonsai, Microsoft has developed Project Moab, a new hardware kit that is available as a simulator in MathWorks, and soon a physical kit for 3D printers. (Developers who don't want to print it themselves can purchase pre-assembled units later in the year.) It is a three-armed robot with a joystick control that tries to balance a ball on a transparent plate attached to a magnet . and to provide an environment for users to learn and experiment with simulations.
Balancing balls is a classic mechanical engineering challenge known as a control problem. Under all conditions, a self-balancing system must learn a control signal to produce the desired final state – i.e. H. A ball that is brought to rest in the middle of the platform. Most classic approaches include differential equations that represent physical quantities and their rates of change. However, Project Moab is trying to find machine learning solutions to the problem.
It's more difficult than it sounds since every ball balancing system needs to be able to generalize – that is, to create a robust control law based on training data. To reach Well For the generalization, a sufficiently extensive set of inputs must be generated during the training phase. Failure to generate a variety of inputs will result in poor performance.
Why build a kit around this problem as opposed to another? Hammond said the Project Moab team wanted to choose a device that engineers and developers could use to learn what steps to take to build an autonomous system. At Moab, developers have to use simulators to model physical systems and include them in a training program. Engineers, many of whom are probably familiar with classic solutions to the problem of ball balancing, must learn to solve it with AI.
"We're giving people more tools in their toolbox to expand the range of problems they can solve," said Hammond. “You can get it very quickly into areas where it would not be easy in the conventional way, such as balancing an egg. The Project Moab system aims to provide a playground where engineers who deal with various problems can learn how to use the tool and simulation models. Once you understand the concepts, you can apply them to your novel use case. "
Project Moab's tutorials are about more than balancing balls. Moab can be taught to catch balls thrown at them after jumping on a table and to rebalance balls that are disturbed after an object hits them like a pencil. It can also learn to balance objects while ensuring that they don't come into contact with obstacles on the plate, much like a self-contained game labyrinth.
Most Moab components – including the plate and arm control actuators – are interchangeable. Developers can install more powerful actuators so that Moab, for example, throws and catches things. The software development kit allows other simulation products and custom simulations to be used to train Moab for more demanding tasks.
Hammond would not rule out future robotics kits for bonsai, but he said that this would largely depend on the community and their response to Moab. "We want the community to be able to experiment and do all sorts of fun, novel things that people haven't thought of before," said Hammond. "Doing open source (a project like this) makes that possible."
Project Bonsai in the real world
SCG is one of the companies that used the Bonsai project to equip their industrial control systems with machine learning. SCG's chemical department created a simulation within the bonsai platform to accelerate the process of optimizing petrochemical sequences to 100,000 simulations per day, each modeling millions of scenarios. Microsoft claims that the fully trained model is capable of developing a sequence in a week, compared to several months previously for a group of experienced engineers.
“Polymers are designed for a specific application. To find out the manufacturing stages, you need to know the mixing, temperature, and other factors, ”said Pall. “The process of creating a polymer manufacturing plan traditionally takes six months because it is done in a simulator in which a human expert guides the simulator and tries one step, finally gets it right, and then the next step continues . Bonsai found a BRAIN that shows solutions for the manufacturability of a particular polymer and then controls machines to make it. "
According to Microsoft, SCG is the first to use a bonsai-trained model in production. In terms of pricing, Bonsai's machine learning component is available to customers free of charge. However, the simulations performed in Azure are billed depending on usage. Companies must purchase a commercial license if they want to use their models in the real world.
Siemens used the Bonsai project for another purpose: the calibration of its CNC machines. Previously, this was a manual process that required an average of 20 to 25 iterative steps over two hours, usually under the supervision of third-party experts. In contrast, the Project Bonsai solution is designed to automate machine calibration in seconds or minutes. According to Siemens, training a model with bonsai made it possible to achieve an accuracy of two micrometers with an average of four to five iterative steps over 13 seconds and an accuracy of less than one micrometer in approximately 10 iterative steps.
"The (Project Bonsai) approach combines AI science and software with the traditional engineering world," said Hammond. "(It enables areas like chemical and mechanical engineering to build smarter, more powerful, and more efficient systems by expanding their own expertise with AI skills."