Composabl holds the key to the future of industrial automation.

Current automation technology, while advanced, lacks in many areas due to its inability to capture and simulate human expertise. The impending loss of institutional knowledge due to retiring skilled workers, coupled with Generation Z’s dwindling interest in manufacturing careers is only broadening this gap, prompting a push to propel industrial automation to the next level.

Composabl’s platform seamlessly integrates traditional automation with AI to create autonomous agents that can simulate human intelligence for real-world tasks. The platform’s success stems from the input of engineers whose hands-on experiences helped shape Composabl’s platform design. Drawing upon their understanding of automation intricacies, engineers can use the platform to build AI agents that merge their organization’s existing technologies and expert operator subject matter expertise with new AI innovations and strategically deploy intelligent autonomous agents on the manufacturing floor.

“Manufacturing fundamentally changes when you put intelligent building blocks in the hands of engineers,” says Kence Anderson, founder and CEO.

Anderson is a visionary leader in autonomous AI, renowned for blending human ingenuity with cutting-edge AI technology, illustrated by the groundbreaking platform he and his team have built.

Next-Generation Automation

Through the Composabl platform, agents can be easily designed to emulate human-like abilities such as perception, learning, strategic thinking, planning and deduction. This allows them to accomplish complex tasks that are next to impossible for a language- based generative AI model and go well beyond today’s automation technologies.

For example, an engineer could use Composabl to enhance the computer numerical control (CNC) manufacturing capability of an oil and gas equipment provider by designing an autonomous agent to control a variety of CNC machines. The agent could use pre-trained classification machine learning models to classify different types of sounds from microphone inputs and employ deep reinforcement learning to become competent in controlling the machine. This creates a modular AI that integrates human decision-making capabilities— specifically perception and learning—and incorporates different strategies. The approach exemplifies how combining traditional automation with advanced AI techniques can enhance manufacturing processes.

“We aim to provide a workbench where engineers can create agents using AI building blocks,” says Anderson.

In the Composabl platform, engineers make use of a no-code, drag-and-drop interface to customize their AI agents. Users can leverage published programs without writing any code themselves, integrating pre-designed components into their systems to build sophisticated agents that orchestrate various skills together. For those who prefer to work with code, it also includes a Python software development kit (SDK), allowing them to publish their custom models and controllers.

Once an agent is built, it is trained using a simulation that engineers or data scientists can publish to the platform. The Composabl Trainer allows the agent to practice in simulation, making tens of millions or even hundreds of millions of decisions as it learns to improve its performance.

Composabl offers multiple ways for users to access the computing power necessary to train at this scale. For initial testing and preparation, training can be conducted on a local laptop. For larger-scale training, users can use the Composabl Training-as-a-Service offering or their own Azure Kubernetes cluster or the like. Each agent is benchmarked for specific goals that guide its training process.

Once an agent has practiced and learned, users can evaluate its performance on the task. Composabl provides an “agent historian,” similar to a factory historian used for tracking factory operations over time. This historian records every action of the agent, along with sensor variables and training scenarios, allowing for comprehensive performance analysis and comparison. Using the historian, users can audit the behavior of different agents and leverage advanced monitoring tools to identify which agents are performing best. Once an agent demonstrates superior performance and meets or exceeds benchmarks, it can be exported for deployment.

Modular AI for Industrial Transformation

In the past seven years, Anderson has personally designed over 200 AI agents for the likes of Fortune 500 industries. Through this experience, he began to recognize a need for modular AI systems that work alongside human operators, such as engineers, and adapt to complex and dynamic environments. Over the course of these projects, a methodology, called the machine teaching methodology, took shape for how to design AI agents with a deep understanding of the specific needs and nuances of the manufacturing processes they are meant to enhance. Anderson dives deeply into the Machine Teaching methodology and his expertise in creating industrial-strength autonomous agents in his book Designing Autonomous AO (O’Reilly).

The Machine Teaching Methodology is the basis for the Composabl platform and Anderson and his team applied their expertise in designing agents to create a platform that allows all engineers to build and train industrial-strength agents at scale.

There are four different approaches to decision-making in the industrial sector; calculation, curating options or solutions, learning by practicing and storing expertise. Composabl’s methodology introduces these approaches in a modular configuration, allowing them to be slotted in or out based on the project requirement. The platform allows users to add different aspects and technologies as they see fit, to design an AI agent for a specific task. For example, they can add deep reinforcement learning to provide learning capabilities, optimization algorithms for planning and a host of audio-visual and prediction models for sensory abilities.

“Our platform can design AI agents similar to the human brain, with flexible components that can perform different decision-making tasks as the need arises,” says Anderson.

The determination of what should be incorporated into a high-performing agent begins with the subject matter expertise of the engineer. The initial stage in creating an effective agent involves breaking down the task into separate skills, creating what is sometimes known as a knowledge graph of the process. This is similar to how a supervising engineer might train their operators, breaking down a complex process into individual steps. This deconstruction of the task into sequential skills essentially outlines a significant portion of the agent’s structure.

In the second stage, the platform facilitates the orchestration of skill sequences. Engineers can pick and choose decision-making approaches to replicate the exact skill configurations known to be effective for a given task.

Our Platform Can Design Ai Agents Similar To The Human Brain, With Flexible Components That Can Perform Different Decision-Making Tasks As The Need Arises

The third and final stage involves selecting the appropriate technology to execute each skill. For instance, a skill that handles safety-critical alerts may be best managed using deterministic, rule-based methods to ensure safety and predictability. Composabl’s methodology integrates these elements by combining the effective organization of agents with the application of suitable technologies.

Shaping the Future of Intelligent Manufacturing

The next step for Composabl involves enhancing its no-code platform and deepening its integration within the industrial manufacturing ecosystem. A significant development under consideration involves integrating large language model (LLM) assistants into its platform. While LLMs are not suitable for direct industrial control tasks, they can support and assist the engineer or expert operator in decision-making.

For example, the assistant could take the form of a chatbot that interacts with users, enabling them to communicate in natural language. It could also act as an executive copilot or assistant, accessing data sources beyond the immediate sensor space, such as news updates, manufacturing trends, and standard operating procedures. By analyzing this external data, the assistant would suggest changes to the agent’s overall behavior or production targets in response to broader context or emerging trends.

Another important focus is expanding the platform’s capability to integrate various decision-making algorithms and technologies. Composabl already allows users to import any Python algorithm into an agent project, but future partnerships will streamline the inclusion of third-party models, simulations, optimization algorithms, and controllers. This will allow their solutions to be incorporated into Composabl’s platform, broadening the range of algorithms and technologies available for use in agent development. Integrating with well-known ML companies and cloud platforms is also an option, to enable data to be imported from a variety of sources for effective agent training.

Composabl is also working on enhancing distribution points where agents can be deployed automatically. This involves developing the capability to push agents directly to control layers, eliminating the need for manual integration. This advance will facilitate the seamless deployment and operation of agents across various control environments.

Ushering in a new era of manufacturing, where automation complements individual expertise, Composabl is bridging the gap between human ingenuity and machine intelligence, poised to transform the manufacturing landscape for the foreseeable future.