Intelligence at the Point of Motion
AI embedment is no longer exceptional at the software layer. Enterprise copilots, service bots, workflow assistants, and generative interfaces are becoming standard features of operating environments rather than strategic differentiators. McKinsey’s 2025 survey found that 78% of organizations use AI in at least one business function, with the strongest reported adoption in IT, marketing and sales, and service operations. Most organizations also still report limited enterprise level bottom line impact, suggesting that adoption is moving faster than structural value capture. Public reporting on AI use generally aggregates enterprise applications under functional categories such as IT, service operations, and knowledge management, meaning platform specific shares for CRM and ERP are often unspecified in published survey data.
The strategic frontier is therefore shifting from software assistance to physical execution. Industrial robotics continues to scale, with global factory robot density reaching 162 robots per 10,000 manufacturing employees in 2023, more than double the level recorded seven years earlier. Professional service robotics is also expanding materially: IFR reports that more than 199,000 professional service robots were sold in 2024, with transportation and logistics accounting for 102,900 units and robot as a service fleets growing 31%. In parallel, self driving systems have moved from concept signaling to measurable live deployment: Waymo reports 170.7 million rider only miles through December 2025 and materially lower injury related crash rates than human driver benchmarks in its operating geographies.
The technical trajectory now points toward a layered model of physical intelligence. Robot foundation models can increasingly connect language, vision, spatial reasoning, and action across different embodiments, but they do not eliminate the need for deterministic control. DeepMind’s Gemini Robotics explicitly preserves the role of low level controllers, while Physical Intelligence’s π0 shows that advanced robot policies still operate through bounded action chunks and staged inference rather than unconstrained end to end motor execution. The more consequential development is that AI is beginning to migrate downward into the electromechanical stack itself. Texas Instruments now positions edge AI motor control microcontrollers as capable of predictive fault detection, adaptive control, and local monitoring of torque, load, and current. The next step in robotics is therefore not just better reasoning at the top of the stack, but more precise intelligence at the point where machines actually move.
For institutions, this changes the axis of competition. The value will not sit solely in model access or interface design. It will sit in system architecture: sensor fusion, simulation, real-time compute, safety validation, update control, actuator design, and operational governance. In robotics and machinery, intelligence without architecture is unstable. Intelligence with architecture becomes infrastructure.
Current State
Most public exposure to AI still occurs at the interface layer. Organizations are using AI across business functions, especially in IT, marketing and sales, service operations, product and service development, and software engineering. McKinsey’s survey data suggest that this is the phase in which AI becomes normalized as software assistance: embedded into workflows, content generation, knowledge retrieval, service interaction, and operational support. But that software centric visibility can obscure a deeper shift already underway in physical systems.
In robotics, adoption is already significant, but more domain constrained. Industrial automation continues to scale because factories reward repeatability, throughput, and measurable ROI. IFR’s 2024 industrial data place global robot density at 162 units per 10,000 employees in manufacturing, while IFR’s 2025 service robot data show transportation and logistics as the largest professional service robot category, with over 102,900 units sold in 2024. These are not speculative indicators. They show that the hardware base for physical AI is already being built at scale.
The same pattern is visible in less structured environments. Amazon’s Vulcan system combines stereo vision, depth sensing, force sensing, and control algorithms to pick from and stow into cluttered fabric storage pods an important threshold because the task depends on controlled contact, not just positional accuracy. In mobility, Waymo’s public safety data show that autonomy can move from demonstration to statistically comparable operations when the sensing, mapping, and safety methodology are sufficiently mature. NHTSA’s 2026 report reinforces that this progression remains inseparable from active research, rulemaking, subsystem testing, and safety oversight.
Key takeaways from the current landscape
Software AI is becoming operationally ordinary. Adoption is broad, but measurable enterprise level value remains uneven, which means embedment alone is no longer a differentiator.
Physical AI is still narrower, but more strategically consequential. Factory robotics, logistics robotics, and bounded road autonomy are already live domains with measurable deployment and safety evidence.
Specific platform level splits are often unspecified. Public surveys tend to report AI by business function rather than by individual CRM, ERP, or chatbot product layer.
Toward Embedded Control and Intelligent Actuation
The technical direction of robotics is increasingly defined by the convergence of semantics, perception, and control. A major inflection point came with the Open X Embodiment effort, which pooled data from 22 robot embodiments and demonstrated more than 500 skills across 150,000 tasks and more than 1 million episodes. The importance of that work was not only scale; it was the proof that robotic competence can be trained across embodiments rather than rebuilt separately for each machine.
That line of development continues in newer robot foundation models. DeepMind’s Gemini Robotics is designed as a vision language action system for direct robotic control, and Gemini Robotics ER extends this into embodied reasoning, spatial understanding, planning, and code generation. DeepMind reports that Gemini Robotics more than doubles performance on its generalization benchmark relative to other state of the art VLA models, while Gemini Robotics ER can interface with existing low level controllers and achieves a 2x–3x success rate improvement over Gemini 2.0 in end to end settings. Crucially, DeepMind is explicit that these models sit within a layered safety architecture rather than replacing motor critical control.
The on device shift makes that architecture more practical. Gemini Robotics On Device is optimized to run locally on the robot itself, without dependence on a data network, and DeepMind states that it can adapt to new tasks with as few as 50 to 100 demonstrations. This matters because many robotic settings are latency sensitive, intermittently connected, or safety sensitive in ways that make remote inference structurally weak. Physical AI cannot scale indefinitely if the machine must wait on the cloud before acting.
Yet even frontier systems still reveal the same engineering truth: intelligence in robotics remains hierarchical. π0, one of the strongest recent examples of general robot control, uses a vision language backbone plus an action expert that produces continuous actions through flow matching. But its own deployment details show that the policy still works through action chunks: for 50 Hz robots, inference runs every 0.5 seconds after 25 actions, and total on board inference is reported at 73 ms on an RTX 4090 class GPU. The model executes those chunks open loop between inference calls. That is not a criticism. It is the architecture. High level learned policies do not replace the timing discipline of low level motion control; they sit above it.
This is where embedded AI in actuators becomes strategically important. Texas Instruments’ 2026 motor control brief is especially revealing because it frames edge AI not as a cloud extension, but as a motor level capability. TI describes local AI models running in motor control MCUs that can proactively monitor torque, load, and current, detect abnormal motor behavior, and help prevent long term damage while also reducing latency and external component count. In other words, the actuator is beginning to become an intelligent subsystem rather than a passive endpoint.
The last step in this trajectory is precise motion control under uncertainty. Variable impedance actuator research remains foundational because it explains why stiff actuation alone is insufficient for dynamic environments, human interaction, or contact rich manipulation. Vanderborght et al. argue that many emerging applications require dynamics that classical stiff actuators do not achieve well, and they position adaptable compliance, variable stiffness, and damping as necessary for safety, motion quality, and energy efficiency. The logic is straightforward: once AI moves into machinery, the decisive layer is not only decision making but how force, timing, damping, compliance, and error tolerance are managed at execution.
Implications for Industry and Labor
In manufacturing, AI enhanced robotics widens the economic boundary of what can be automated. The opportunity is not simply faster repetition, but handling more variability without bespoke reprogramming every time conditions change. That is why contact aware manipulation matters so much: it expands automation into tasks that were previously too cluttered, too deformable, or too variable for rigid industrial logic.
In logistics, the signal is already clearer. IFR’s service robot data show transportation and logistics as the dominant professional application class, while Amazon’s Vulcan offers a live example of why: warehouse environments produce repeatable task families, real economic pressure, and abundant operational data, making them the natural proving ground for perception, manipulation, and fleet optimization. The more AI improves exception handling, grasp planning, and local adaptation, the more logistics becomes a systems problem rather than a labor scaling problem.
In mobility, the implication is narrower but highly significant. Autonomous driving still remains bounded by geography, regulation, and safety case design, but it is no longer only a demonstration category. Waymo’s safety data and NHTSA’s ongoing work on ADS rulemaking, subsystem evaluation, simulation, crashworthiness, and standards modernization indicate that bounded autonomy is becoming institutionally real before universal autonomy becomes technically or politically acceptable.
For labor, the most credible short run outcome is redesign rather than total replacement. McKinsey Global Institute estimates that currently demonstrated technologies could, in theory, automate activities accounting for about 57% of US work hours today, while more than 70% of skills sought by employers are used in both automatable and non automatable work. This suggests a shift in task composition rather than a clean subtraction of human relevance. As machine systems take on more execution, human work moves further toward supervision, exception handling, interpretation, maintenance, orchestration, and safety judgment.
Engineering Constraints That Still Matter
The final mile remains difficult because physical systems must satisfy more than capability demonstrations. First, sensing must become more multimodal and more reliable. Amazon’s Vulcan uses force sensing and depth perception because vision alone is not enough for contact rich execution. In softer and more complex morphologies, the challenge grows: a 2024 Nature Communications paper demonstrated a soft continuum robot with 1.4% average shape estimation error, contact detection within 0.4 seconds, and contact direction error below 10°, but also explicitly identifies high level perception action loops as one of the major unresolved challenges in soft robotics.
Second, control loops and inference loops still run on different clocks. NHTSA breaks ADS functionality into perception, decision/path planning, and execution/control, and that decomposition remains useful far beyond road autonomy. π0’s chunked execution and DeepMind’s continued reliance on low level controllers both point to the same conclusion: large learned models can improve planning and generalization, but deterministic motion control still has to be engineered within strict time and safety bounds.
Third, safety and verification become harder as systems become more adaptive. NIST’s AI RMF structures risk management around govern, map, measure, and manage, and explicitly notes that real world operational risks can differ from pre deployment laboratory measurements. ISO 10218-1:2025 likewise emphasizes inherently safe design, risk reduction, and information for use at the robot level before full system integration. In physical AI, model performance is insufficient without lifecycle assurance. Trust must be designed into the system.
Fourth, energy, materials, and actuator design remain central. NIST notes energy and environmental implications associated with resource heavy AI computing, while TI’s edge motor control approach is explicitly aimed at lower latency, lower overhead local execution. At the machine level, this means AI progress is constrained not only by models, but by thermal budgets, package sizes, component reliability, and actuation physics. The next leap in robotics will come from co-design across compute, sensing, and mechanics not from software alone.
Timelines and Scenarios
The most plausible near term path over the next one to three years is accelerated deployment in bounded environments: warehousing, machine tending, visual inspection, predictive maintenance, guided mobility, and selected industrial handling tasks. The operative shift will be from fixed automation to systems that can absorb a higher degree of variability without full re-engineering.
The most reasonable medium term path over three to seven years is broader adoption of layered physical-AI stacks: foundation model planning at the top, local action generation in the middle, and machine level embedded intelligence closer to motors, joints, and sensing hardware. More robots will operate semi autonomously across variable workflows, but most production systems will remain supervised, bounded, and heavily instrumented rather than fully free form.
The more credible long term path over seven to fifteen years is not a simple proliferation of humanoid demos, but the rise of self optimizing physical systems: machines that estimate wear, detect fault states, adapt control locally, and improve across fleets through governed update loops. Whether those systems become widespread will depend less on headline model progress than on verification methods, safety acceptance, actuator maturity, and institutional willingness to redesign workflows around physical intelligence rather than retrofit it onto old architectures.
Enterprise AI and Work Systems
McKinsey & Company (2025) The state of AI: How organizations are rewiring to capture value.
McKinsey Global Institute (2025) Agents, robots, and us: Skill partnerships in the age of AI.
Industrial Robotics, Logistics and Mobility
International Federation of Robotics (2025) World Robotics 2025 report – Service Robots.
Amazon Science (2025) How Amazon’s Vulcan robots use touch to plan and execute motions.
Robot Foundation Models and Embedded Control
Google DeepMind (2023) Scaling up learning across many different robot types.
Google DeepMind (2025a) Gemini Robotics brings AI into the physical world.
Google DeepMind (2025b) Gemini Robotics On-Device brings AI to local robotic devices.
Physical Intelligence (2024) π0: A Vision-Language-Action Flow Model for General Robot Control.
