The following are the key considerations that GMR has learned from deploying robotic systems in real factories. These considerations shape the development of Physical AI for manufacturing applications and form the basis of an ensemble approach:
- Diversity of Information Sources entails Multi-level Abstractions: Manufacturing information is inherently diverse, spanning multiple formats and levels of abstraction. This information includes textual and graphical standard operating procedures, mathematical equations and computational simulation models, real-time machine data, and process videos. To be truly effective, AI applications in manufacturing must be capable of synthesizing and leveraging all of these disparate data forms. Moreover, there is no single information abstraction level that is suitable for making different decisions. Therefore, multiple different levels of abstractions need to be maintained by AI models.
- Multi-Level Decision Making requires Hierarchical Intelligence: Decision-making in manufacturing is often at multiple levels, spanning macro-level planning to micro-level physical control. Macro-level decisions govern task decomposition and scheduling. These dictate subsequent process-level choices regarding tooling and skill allocation. At the execution layer, real-time control decisions regulate tool motion and force regulation. High-mix manufacturing uses sensors to identify objects and build part models. This also requires decisions related to sensor positioning. Decisions are also needed related to part quality assessment. To be useful in manufacturing applications, AI needs to be able to support hierarchical decision making.
- Multi-Modal Sensing needs Task-Aware Fusion: Manufacturing applications often rely on multiple sensing modalities, including force, RGB vision, 3D vision, vibration, thermal imaging, ultrasonic, eddy current, machine telemetry, and more to monitor both the workspace and the process. These sensors operate at different frequencies, measure different physical phenomena, and vary in relevance depending on the task being performed. Therefore, sensor fusion cannot be a generic perception problem. It must be task-aware, combining the right modalities at the right time to estimate the system state, assess process quality, and guide the next action. Moreover, we also need AI that monitors sensors and detects when sensors are not functioning well. For example, excessive dust on a camera may interfere with its functioning. This requires an ensemble of modality-specific perception models, fusion models, task-state estimators, and sensor-monitoring modules rather than a single universal sensing model.
- Explainability requires Traceable Decision Logic: Manufacturing environments are regulated and operating under strict liability laws. Therefore, in manufacturing, deploying an AI model is different from deploying one in a consumer app. If a movie recommendation model fails, a user clicks the wrong video; if an AI model fails in manufacturing, it can result in destroying a part costing tens of thousands of dollars, days of factory downtime, or severe operator injury. If the system recommended actions do not match user expectations, then the system should be able to explain the rationale used to select actions.
- Hybrid Computing Architecture for Distributed Execution: Performing all AI computation in the cloud is not viable in the context of Physical AI. The system should be designed to ensure that computation that is sensitive to network latency can be performed on the edge, hence this demands distributed execution in a factory. The use of edge computing may also be needed to support applications where sensor data cannot be transmitted to clouds due to privacy/security reasons.
- Rapid Upgrades require Modular Adaptation: The performance of a physical system may change over time because of wear and/or updates to physical components. This may require refinement of AI to ensure that it can keep up with system evolution over time. Therefore, the physical AI system needs to be designed to ensure that it can be updated with minimal disruption to the system operation.
Pretrained Components Enable Rapid Cell-Level AI Configuration: Manufacturing robotic cells can come in many different configurations based on the size of the part being processed and the process being performed (e.g., sanding or blasting). Different cells may include robots with different capabilities (mobile platform mounted robot vs. gantry mounted robot), different types of sensors (depth camera vs thermal camera), and different tools (orbital sanders vs blasting nozzle). Therefore, generalization in manufacturing requires reusable AI capabilities that can transfer across cells while still being configured to the specific sensing, actuation, process, and work environment of each deployment. The AI for the system needs to be quickly synthesized from pretrained components so each cell can benefit from common learned capabilities without requiring a new system to be built from scratch.
Manufacturing requires deep expertise across many different functions: industrial engineering, process engineering, deployment, operations, prognostics and health management, and sustainment among others. Dedicated domain-specialist AI-powered agents are needed for each of these functions. What makes these agents truly domain-specialists, rather than general-purpose automation, is that each is built to reason and act within the full context of its function.
Different Manufacturing Skills Require Different Learning Paradigms: There is no single AI method that is best suited for every physical skill required in manufacturing. Some skills are most effectively learned through imitation, where expert demonstrations provide a strong starting point. Other skills require reinforcement learning, where the system improves through trial, feedback, and optimization. In certain cases, traditional model predictive control remains the right tool because the physics are sufficiently well understood and real-time constraint satisfaction is critical. For high-level planning, foundation models can help reason over language, images, and prior knowledge. However, when execution carries significant risk, world models are needed to simulate, validate, and de-risk proposed actions before they are executed on the physical system. This has parallels with how humans develop expertise. Skilled operators do not rely on a single mode of intelligence. They imitate experienced workers, practice and refine their actions through feedback, rely on muscle memory for familiar tasks, reason explicitly when conditions change, and mentally simulate outcomes before taking risky actions. Manufacturing AI needs the same kind of flexibility. The right architecture must be able to combine imitation, exploration, control, reasoning, and simulation, using the right learning paradigm for the right skill and the right level of decision-making.
Key Features of Ensemble Approach: We are developing tools to rapidly configure new AI-powered solutions based on new materials and customer requirements. The architecture of the underlying Physical AI used to power robotic cells is based on an ensemble approach. Key features of this approach include:
- Select the right functional decomposition to achieve the best tradeoff between modularity and performance: Our systems use the right functional decomposition to ensure that it is able to achieve the desired tradeoff in performance and modularity. This ensures the ease of upgrades and enables rapid generation of a new AI system to match a new hardware configuration.
- Select the right AI approach for each functional block: As explained earlier, many different AI approaches exist to acquire new skills. Each approach has its pros and cons. It is unlikely that a single approach will suffice to deliver the desired performance. Therefore, we carefully select the right AI approach for each functional block in our system.
- Exploit known models: We use AI approaches that exploit the known models and use a data-driven approach to augment the known models (and existing knowledge) based on experimental data to fill the missing model gaps.
- Develop clear and easy to understand specifications for interfaces between functional components: Ensuring ease of the failure diagnosis and enabling explainability requires us to use clear and easy to understand specifications for interfaces.
