1. Introduction
The emergence of embodied artificial intelligence (AI), where intelligence is exemplified through physical interaction with the environment [
1], [
2], has permeated through a variety of industries. In manufacturing, the evolution of integrating mechanical, electrical, and computational elements has long driven innovations [
3] and is foundational to the emergence of embodied AI. This evolution is fueled by the widespread deployment of sensors, digitization of physical processes, and the increasing availability of data, leading to intelligent systems that are contextual, goal-directed, and physically embedded [
4].
Embodied AI represents a paradigm shift in the design and construction of intelligent physical systems, where emphasis is placed on the integration of physical interaction and situational awareness of such systems to facilitate adaptive decision-making. This embodiment is realized through an interplay of sensing, control, and actuation, where AI not only interprets data but directly interacts with physical processes. Enabled by advances in deep learning (DL), sensing technologies, and computational infrastructure, embodied AI systems permit new avenues for a broad range of manufacturing applications by building upon a series of transformative capabilities that span semantic data inference, learning-based control, and generative optimization of actuation hardware design. This paper presents an overview of recent advances to enable embodied AI for manufacturing and outlines promising research directions.
As AI continues to impact all aspects of manufacturing [
5], its evolution from a passive analytical tool to an active, embodied agent represents a significant leap in capability. In this new paradigm, machines no longer only interpret data or execute preprogrammed commands but actively adapt to their environment through learned experiences. Traditional AI applications in manufacturing have primarily emphasized data interpretation and decision support. The growing demand for autonomy and adaptability calls for systems that can directly sense, act, and learn through physical interactions. This transition from analytical to embodied AI is driven by the imperative to close the perception-action loop, where intelligence emerges from continuous engagement with the environment, enabling manufacturing systems to evolve from reactive instruments into proactive, self-optimizing agents, thereby ushering in a fundamental paradigm shift in how manufacturing operations are conceived and executed.
Underpinning this capability are the cornerstone technologies that have emerged recently: ① data-driven sensing that enables machines to infer semantic and actionable insights from raw measurements; ② data learning-based control that allows systems to adapt to dynamic and uncertain environments; and ③ generative design approaches that optimize physical embodiment for specific tasks. This paper is organized around these three cornerstones, and outlines the emerging topics such as digital twins, large language models (LLMs), and agentic AI to further push the boundaries of digitization in the physical world, foundation models, and agent-based decision-making.
As illustrated in
Fig. 1, embodied AI in manufacturing operates as a closed loop linking foundation models and LLMs with decision-making agents and physical systems through three interconnected processes: sensing, control, and actuation. Data and feedback circulate among these components, allowing the system to perceive its environment, update control strategies, act upon the physical world, and refine its behavior through continuous digital-twin simulation. This cycle embodies the integration of learning, reasoning, and interaction that defines intelligent and autonomous manufacturing.
2. Data-driven information inference for situational awareness
To enable embodied intelligence, physical interaction must be informed by situational awareness derived from sensing data. Conventional signal transduction methods capture physical phenomena, but do not directly support decision-making. Data-driven methods close this gap by mapping raw sensor data to high-level latent features [
6]. These techniques offer the potential to convert noisy and complex measurements into robust, task-aware insights, hence fundamentally shifting the paradigm from sensing to understanding. At the heart of data-driven sensing are a series of DL-based techniques and network architectures that are tailored to specific data types enable inference of object properties, human locations, or process states that are hidden otherwise [
7].
2.1. Semantic segmentation
One representative technique is the deep encoder-decoder for semantic segmentation of images, such as material surface identification in robot-assisted processes, where the encoder extracts low-level features and distills them into high-level representations of surface patterns. The decoder then reconstructs pixel-level segmentation masks. This ability to parse the scene semantically from visual data forms the foundation for downstream control systems to perform corrective actions, such as weld seam removal [
8].
2.2. Depth estimation
In collaborative robotics, depth estimation using red-green-blue (RGB) cameras allows robots to locate objects and human co-workers in dynamic workspaces. Tackling the challenges facing traditional methods such as triangulation, learning-based depth estimation complements the analytical foundation of these classical methods with data-driven mapping to learn to associate depth maps with RGB images, overcoming limitations such as the scenario of both camera and target are in motion [
9]. Such methods remove the need for dedicated depth sensors, increasing the scalability of embodied AI systems under constrained resources.
2.3. 3D environment reconstruction
Rapid three-dimensional (3D) environment digitization is crucial to enable dynamic and safe interactions within embodied AI systems. The emergence of data-driven neural radiance fields (NeRF) tackles this problem by inferring spatial distributions of its color and density from multi-view images [
10], which allows reconstruction of 3D environments containing transparent or complex objects without explicit camera calibration. This capability is crucial for tasks such as robot path planning. NeRF also opens new pathways for learning-based grasp synthesis and geometric reasoning, and serves as a bridge between perception and simulation.
2.4. Tactile interpretation
As contact becomes commonplace for embodied AI systems, tactile sensors are envisioned to play an increasingly important role. Complemented by DL’s capability to learn transformations, complex relationships between raw sensor readings and semantically interpretable tactile information can be obtained. As an example, conductive elastomers can transduce contact force into resistance changes, which can be transformed into contact types (e.g., point, line, and surface) by DL models to support real-time grasp adjustments by robotic grippers [
11].
2.5. Proprioception
In soft robotics, knowing the exact posture or proprioception is of critical importance. With DL, this can be achieved by mapping voltage changes in embedded sensor arrays to 3D configurations of soft robotic arms, thereby overcoming the need for external sensing systems [
12]. As soft robots become increasingly relevant for delicate or bio-inspired tasks, the capability to internalize the robot’s own configuration is key to achieving safe, precise, and adaptive interactions with variable environments.
3. Learning-based control for adaptive and safe decision-making
Learning-based control lies at the core of embodied AI, enabling systems to respond dynamically to complex environmental feedback and physical interaction in a manner that is beyond the capability of classical control algorithms. Whether stabilizing melt pool behavior in additive manufacturing or optimizing robotic grasp in uncertain settings, learning-based control strategies translate perception into informed action and form a critical bridge between sensing and actuation.
3.1. Model predictive control (MPC)
MPC solves constrained optimization problems over short prediction horizons by anticipating system behavior using state-space models. Due to MPC’s reliance on physics-based interpretation, the method is increasingly empowered by data-driven approaches that enable control of nonlinear, highly variable systems [
13]. Physics-informed machine learning further advances this synergy by embedding physics-based governing equations, yielding models that are interpretable and generalizable. As manufacturing systems grow in complexity, AI-augmented MPC offers a principled yet flexible framework to ensure performance and robustness.
3.2. Model-free reinforcement learning (RL)
RL presents an alternative paradigm for control, where agents learn optimal actions through direct interaction with the environment [
14]. Unlike MPC, RL does not require an explicit process model, making it attractive for applications where modeling is infeasible or costly. Examples such as control of deep drawing and robotic grasping highlight RL’s strengths—its ability to learn directly from experience, accommodate sparse and delayed rewards, and discover novel strategies, equipping embodied AI systems with the adaptability and autonomy needed to operate in evolving manufacturing environments.
Taken together, MPC and RL provide complementary strategies for embodied control—one offering structured, constraint-aware optimization, and the other delivering flexibility and self-improvement through exploration, respectively. As an example, safe RL can be achieved through integration of an MPC-based safety filter and an RL-based controller, where the objective of the safety filter is to find, commonly through optimization, a control command that is the closest to the initial command calculated by RL while staying within predefined constraints (
Fig. 2). The integration of these methods enables a new class of systems that are not only adaptive but also physically grounded, through the explicit satisfaction of physical constraints ensured by MPC and the incorporation of physics-informed machine learning within RL, where governing equations describing the underlying process dynamics are embedded to guide policy optimization and ensure physically consistent control.
4. Generative design of actuation for physical embodiment
For embodied AI systems to interact effectively with their environment, the design of their physical components, particularly actuation mechanisms, is critical. However, discovering optimal physical designs of such systems is challenging. The design space is often high-dimensional, involving constraints that are difficult to encode analytically. Traditional trial-and-error and expert-driven processes are time-consuming and suboptimal when dealing with such complexity. The interdependence between geometry, material properties, control strategy, and task requirements further complicates the search for high-performance designs.
End-to-end generative learning frameworks can represent design generation as a self-supervised optimization task, offering a powerful tool to navigate the complexity of physical designs. Specifically, a generative DL model proposes design candidates, which are evaluated by a downstream simulation or learned performance proxy. Feedback quantifying how well a design performs guides the generative process to refine its future proposals (
Fig. 3). This closed-loop, data-driven methodology transforms hardware design into an iterative learning process, thus eliminating the need for hand-crafted features or exhaustive search [
15] and demonstrating the potential of embodied AI to
co-evolve hardware alongside sensing and control.
While demonstrated here in the context of generative design, the same self-supervised paradigm can be extended to adaptive and safe decision-making. By leveraging high-fidelity simulation, controllers such as MPC- or RL-based agents can autonomously learn and validate control policies in virtual environments prior to deployment on physical systems, thereby ensuring both safety and adaptability. This concept naturally aligns with recent developments in multimodal sensing, LLM-based reasoning, and digital-twin integration. Collectively, these advances form the foundational infrastructure for achieving self-supervised learning in embodied manufacturing systems.
5. Conclusions and outlook
Embodied AI, where intelligence arises from direct process interaction, has begun reshaping manufacturing. This paper highlights how recent development in data-driven sensing, learning-based control, and generative design provide the foundation to enable systems that interpret, decide, and act in adaptive ways. While initial progress has been made, many promising research topics remain to be explored.
5.1. Data-driven co-design
In current methods, sensing, control, and actuation are often optimized independently. Integrating these into a unified, end-to-end trainable pipeline will allow embodied AI systems to optimize for performance, adaptability, and efficiency simultaneously [
16].
Beyond robotic systems, data-driven co-design principles could be extended to reconfigurable manufacturing platforms where sensors, actuators, and AI inference modules are modular and dynamically linked. Such adaptive architectures would allow the same hardware to morph into different configurations depending on the task at hand [
17]. A recent study demonstrated a digital-thread based framework in which production modules are designed, configured, and executed in a flexible manufacturing line using a co-design approach that couples product design, sensor/actuator layout, and process execution [
18]. This paradigm provides a pathway toward embodied AI in manufacturing, where physical systems and AI models co-evolve—perception modules inform control design, control decisions influence actuation and layout, and real-world feedback continuously update digital models. By embedding this closed co-design loop, embodied AI can enable highly flexible and adaptive manufacturing, from autonomous assembly cells to reconfigurable large-scale production systems.
5.2 Foundation for digital twins
Digital twins require real-time synchronization between a digital model and a physical system [
19], [
20], [
21]. Embodied AI systems form the computational foundation of digital twins by enabling monitoring, diagnostics, and closed-loop optimization—data-driven system identification and predictive modeling serve to update the digital twin, while control modules translate situational awareness into executable feedback. In turn, the digital twin can guide embodied agents through a simulated improvement loop, enhanced using generative AI and its capability to synthesize high-fidelity environments for the agent to interact with and adapt to its physical twin, thereby closing the sim-to-real gap. Such integration facilitates the translation of embodied systems developed within digital twins into real-world scenarios, further amplifying system resilience and supporting lifecycle analysis in manufacturing.
5.3. LLMs and agentic AI
Recent advances in LLMs and agentic AI suggest new interaction modalities between humans and embodied systems. LLMs, often referred to as a new class of foundation, generalist models, can serve as natural language interfaces for a wide range of general tasks for embodied AI [
22], while agent-based frameworks can orchestrate sequences of actions, enabling high-level planning [
23]. Depending on the system hierarchy, an LLM may serve as either a high-level commander that plans and orchestrates the actions of embodied agents, or as an active player directly engaged in decision-making through close integration with perception and control modules.
For example, AI agents that are capable of planning, memory, and coordination can abstract low-level sensory-motor skills and dynamically invoke learned policies to create flexible actions. Moreover, LLMs trained with domain knowledge can potentially assist in translating design and control specifications into machine-readable plans, bridging the gap between engineering and AI execution. Recent developments include an LLM-based manufacturing process planning system that converts informal task instructions into structured, executable workflows for flexible manufacturing equipment, thus coordinating devices and modules in the context of Industry 5.0 [
24] and creating a hardware platform embodied with AI at the device level. In another case, LLM-agent integration with a flexible modular production line enabled planning and control of previously undefined tasks via agent orchestration [
25].
Recent advances in vision-language models and multimodal generation further strengthen this framework by grounding linguistic reasoning in visual and sensory contexts, enabling embodied AI systems to interpret human instructions and act coherently across modalities.
5.4. Multimodal embodied AI
The evolution of embodied AI naturally progresses toward multimodal sensory integration, where visual, auditory, tactile, and linguistic data are fused to achieve deeper and more comprehensive situational awareness. In manufacturing, such multimodal fusion enables robots to correlate acoustic signatures of processes, visual states, and contact forces, thereby enhancing the reliability and safety of perception-driven decision-making while facilitating intuitive human-robot collaboration through natural interactions.
A recent example is the vision-language-model-driven robotic manipulation system, where scene understanding and object handling are guided through integrated vision and language reasoning. Extending this paradigm to include additional modalities such as voice and force sensing can further enhance adaptability, transparency, and resilience of embodied systems in complex manufacturing environments characterized by uncertainty and disturbances [
26].
5.5. Human-centric and resilient AI
As manufacturing environments are human-centric [
27], embodied AI systems must be designed for safe coexistence, mutual awareness, and continual collaboration with human operators. Humans act as an integral component of the loop—providing corrective feedback, contextual judgment, and demonstrative guidance to support adaptive learning and decision refinement [
28], [
29]. Achieving effective integration requires aligning explainability with AI-driven control decisions, adaptive safety envelopes, and cooperative learning. These are the key challenges that define the path toward truly collaborative intelligence in manufacturing.
Furthermore, resilience must be inherently built into the system to enable recovery from anomalies or sensor failures without significant degradation of its functional performance. Such resilience increasingly depends on the synergistic interplay between human domain knowledge and the adaptive capabilities of embodied systems, where human feedback accelerates recovery and fosters transparent, trustworthy operation over prolonged deployment.
5.6. Ethics and governance
With increased autonomy comes the need for ethical frameworks and governance structures. Transparent reporting of AI behavior, interpretability of decisions, and alignment with societal goals are essential to trustworthy deployment in manufacturing [
5].
Continued interdisciplinary collaboration between the AI and manufacturing communities will be key to advancing embodied intelligence. By integrating perception, decision-making, and actuation into a cohesive unit, embodied AI offers a pathway toward adaptive and agentic intelligent systems for manufacturing in Industry 5.0.
PDF
(715KB)
