Frontier Technologies in Industrial Operations 2025

2.2 Embodied AI – igniting a new era in robotics AI is not only transforming software but also automating physical workflows. Embodied AI integrates AI into physical systems such as robots, allowing them to perceive and interact with their environment through dynamic and complex movements. The agents see the world via sensors (for example, cameras, radar, lidar and microphones) and execute actions through actuators such as advanced grippers. Applied to industrial operations, these agents enhance the capabilities of existing robotic systems, enabling more sophisticated automation. By doing so, they expand the automation scope, overcoming traditional challenges such as those associated with handling unstructured environments or manipulating unstable objects. Three types of robotic systems FIGURE 2 Robotic software improvementRobotic hardware improvementRobot capabilities enabled by embodied AI Rule-based robotics Coding Performance and reliabilityTraining-based robotics Training Task versatility and flexibility Situation adaptability Manipulation dexterityContext-based robotics Zero-shot learning General understanding and task execution Human-like dexterity and low-level control Universal robotic embodiment Source: Boston Consulting Group. Notably, three types of robotic systems have emerged: rule-based, training-based and context- based (Figure 2). This evolution has been driven by improvements in both robot hardware and software. The hardware is becoming more capable, reliable and flexible. At the same time, the software is advancing, with improvements in foundation models and technologies (such as reinforcement learning).8 A five-fingered robotic hand with 24° of freedom can perform complex tasks with an unprecedented level of dexterity.9 This is made possible by the various data sources that can be harnessed to train AI-enabled robots: –Real robot data: This data is collected from the robot motion controllers and can also be generated by human-guided robot teleoperation. Although real robot data is the most accurate, it is limited because it can only be gathered from deployed robot fleets. –Synthetic robot data: This data is created in simulated physics-based environments and is available in infinite supply. While any scenario can be simulated, a simulation-to-reality gap is expected to remain due to the diversity of the real world. This means real robot data will still be necessary for validation. For example, Foxconn trains robots in its virtual factory, using digital twins to generate synthetic data for model training and to teach robotic arms how to see, grasp and move objects.10 –Internet-scale human data: Online data, including human videos, is highly diverse, and equips AI with a foundation for understanding the world. It provides valuable information on how humans interact with objects and how objects behave. Imitation learning allows the latest models to learn these skills by mimicking human actions. The discussion of the three robot types enabled by embodied AI centres on these software advancements, which harness advanced datasets: Rule-based robotics: Beginning in the 1960s, industrial robots operated under rule-based systems, following “if… then” instructions that were manually coded by experienced robotic engineers. Complex Embodied AI integrates AI into physical systems such as robots, allowing them to perceive and interact with their environment through dynamic and complex movements. Frontier Technologies in Industrial Operations 15