Generalized AI: An approach that uses the same ML model for applications in multiple domains. The goal is to cover as much of the input space as possible, and the belief is that training on data from one domain can help performance in another.
Specialized AI: An approach wherein each domain or application has its own specific model. The goal is to learn to do one thing expertly, and the belief is that by focusing on one section of the input space, models can be smaller, faster and require less data.
Adaptable AI: An in-between approach that recognizes general models may not perform well on rare or domain-specific inputs and that specialized models are inflexible. The goal is to incorporate real-time feedback to modify the model’s area of specialization in response to changing inputs.
General models claim broad applicability but may not perform well on truly unique aspects of a particular application. In addition, as data and power-hungry technologies, these models can be quite expensive to build and maintain. Meanwhile, focused, adaptable models that self-specialize are immediately effective and far more affordable. That said, there is room for both approaches to influence the future of automation and robotics.
Consider this: A warehouse requires a robot capable of de-palletizing rainbow pallets. This robot improves warehouse efficiency by relying on AI to identify new package shapes and adjust its pick points accordingly. However, it performs the same rote programming tasks most of the time. In other words, once the AI has driven the robot to the correct pick position, the rest of the process — including placing the carton and returning to the pallet for the next cycle — is simple and repetitive. Thus, the focus is on solving the vision problem - How does the system know where and what to grab?
The promise of generalized AI
Generalized AI supporters argue that models derived from specialized training sets are too limited compared to those based on more generalized data sets, also known as foundational models. Furthermore, these experts believe in the power of cross-training, where a model trained on data from one application sees a performance boost when performing a different activity, enabling robots to generate outputs for novel inputs based on broad understandings of the world.
Generalized AI is indeed an exciting possibility — but for warehouse logistics, it remains a possibility. Models capable of generalizing over all possible inputs haven’t yet been achieved, so we simply cannot deploy one on the warehouse floor.
To achieve and validate full generalization, a system must train, test and collect data over the entire space of possible inputs a system might see — an expensive and time-consuming proposition that is only as good as the forecasting of future variability. In contrast, adaptable models only require data from — and only spend time training and testing on — scenarios that actually occur in the application at hand.
Furthermore, general models assume that the proper output for a given input is the same in all cases. That is, when you see X, you must do Y. However, this is not the case across all applications or even all instantiations of a particular application. For example, if the system sees a tape gun, is that something to pick up and place on the outbound conveyor, or is it a tool left there by maintenance personnel that should be avoided? It depends, and in different applications, each reaction could be a correct response. For general models, this results in the so-called 'multimap' problem, where hidden information is necessary to differentiate between possible outputs. In practice, a solution is often achieved by adding additional, human-configured inputs into the system, effectively partitioning the input space in much the same way as multiple specialized models already do.
The practicality of adaptable AI
Given the state of general models today, we believe that models relying on continuous adaptation are better suited for warehouses and logistics operations. Robots running such systems can be set in motion on day one and contribute meaningfully to an e-tailer or 3PL’s ability to meet delivery windows. Simply put, these systems are focused on doing what needs to be done and are thus more skilled and adept at their jobs. As for generalized AI models, think of those as “jacks of all trades” — they have broad experiences and can be helpful in many applications but are not truly great at any.
A warehouse environment is always in flux, the key reason to consider adaptation AI. In these environments, change is constant, even after 1 billion picks in production, Plus One systems continue to encounter new scenarios every day, exactly why we have Yonder. Yonder, our human-in-the-loop approach for adaptable machine learning, routinely moves those scenarios from exceptional to expected. Human-in-the-loop supervision guarantees that a system only learns the right lessons, enabling it to handle similar inputs autonomously in the future while ensuring that exceptions never result in inefficiencies or delays.
Looking toward the future of robotics
Advancements in one field of robotics benefit us all, from assembly line workers seeking more fulfilling work to warehouse leaders attempting to meet tight shipping deadlines. Therefore, we welcome the day in the far future when generalized AI can bring value to unstructured tasks and warehouse automation.
But in terms of today, in light of the variability in warehouses here and now, adaptable AI is not only superior — it’s the only viable option.
To learn more about the viability of adaptable AI in warehouse automation in your warehouse, contact us at https://www.plusonerobotics.com/contact.