A different kind of AI is currently being developed, the so-called "Embodied AI." It refers to agents that have a body and support physical interaction, such as intelligent service robots, self-driving cars, etc.
Embodied AI robots can interact with the environment, plan, make decisions, act and perform tasks like humans. For example, the robot unit is tasked with sanding the upper surface of a part placed in the unit to achieve the desired surface finish. Embodied AI is able to use sensors to monitor the state of the unit and generate instructions for the robot to perform tasks.
Digital AI and embodied AI share some similarities and utilize many underlying technologies. However, understanding the differences between these two types of AI is critical to successfully applying digital AI methods to specific AI applications.
The risk profile of embodied AI applications is often fundamentally different from that of digital AI applications. If digital AI tools are 99 percent accurate, it could dramatically improve human productivity in many applications.
In contrast, due to the risks of industrial applications, the accuracy requirements for specific AI systems often vary widely.
The main risks come from two aspects: the probability of error and the consequences of error. When the consequences of making a mistake are not serious, a higher probability of error can be tolerated. This is why a 1% error probability is acceptable in many digital AI applications.
Conversely, many embodied AI applications require error probabilities better than one in a million. Using a purely data-driven approach to reduce the probability of errors requires a lot of data. In most cases, the demand for data is growing exponentially. Unfortunately, the cost of getting data from physical systems is high. Therefore, a different approach needs to be followed when dealing with embodied AI applications.
To meet the above requirements, embodied AI for manufacturing applications should have the following characteristics:
Training with limited data: The embodied AI can be trained with limited data generated from physics experiments first.
Can be assembled from pre-trained modular components: physical systems can have multiple configurations to support their intended needs. For example, depending on the process being performed (such as sanding or sandblasting), the manufacturing robot unit can be in many different configurations. Different units may include robots with different functions (such as mobile platform mounting robots or gantry mounting robots), sensor types (such as depth cameras or thermal imagers), and tools (such as orbital sanders or sandblasting nozzles).
As a result, developing universal embodied AI that works out of the box for all manufacturing applications may not perform very well. The AI of the system needs to be synthesized quickly from modular components to match the sensing and driving capabilities of the specific system and work environment.
Can be adapted to new data or context: As new data becomes available during system deployment, it should be possible to use this data to improve AI performance. AI should be able to adapt autonomously to new environments or tasks with minimal human supervision.
Easy to upgrade: Over time, the performance of the physical system may change due to wear and tear or updates to the physical components. This may require improvements to the AI to ensure it can keep up with the evolution of the system. Therefore, an embodied AI system needs to be designed to ensure that it can be upgraded with minimal disruption to the system's operation.
Risk-based recommendations for action: The system should be able to estimate its confidence in the proposed action. When confidence is low, the system should conduct a risk analysis and analyze the consequences of failure. If the risk is too high, the system should seek help from human experts.
Interpretability: If the system suggests an action that does not meet the user's expectations, the system should be able to explain the reasons used to select the action.
Distributed architecture that supports partitioning of computing between edge and cloud: In embodied AI application scenarios, it is not possible to perform all AI computing in the cloud. The design of the system should ensure that network latency-sensitive calculations can be performed at the edge.
In the field of digital AI, we are seeing great success with large end-to-end learning models such as LLM. These models thrive on huge amounts of data. However, they do not possess many of the characteristics of embodied AI mentioned above.
Embodied AI should be viewed as a complex system involving interactions between multiple AI components. Having the right system architecture in embodied AI is one of the keys to successful manufacturing applications. This enables you to take advantage of the latest advances in AI and meet the demanding requirements of manufacturing applications. Therefore, modern systems engineering methods are needed to design embodied AI for manufacturing applications.
