From automation to physical intelligence: why robotics is becoming an operating model and data question

Robotics, or “Physical AI”, is entering a new phase. The question is no longer simply whether robots can perform more tasks. It is whether organisations can redesign work so robots create reliable, safe and economic value at scale — and whether they can capture the real-world data needed to improve those systems over time. 

That is why physical robotics is no longer simply a hardware story. It is becoming an operating model story, and increasingly, a data story. 

By physical intelligence, we mean AI-enabled systems that can perceive, reason and act in the physical world. But their commercial value depends on much more than the machine itself. It depends on the workflow, safety model, data layer, human-machine hand-offs and operational system around the robot. 

The market momentum is already significant. More than 542,000 industrial robots were installed globally in 2024, bringing the operational installed base to 4.66 million systems worldwide. The International Federation of Robotics expects annual installations to exceed 700,000 by 2028, with China accounting for 54% of global deployments in 2024.   

This is no longer a niche technology trend. It is an infrastructure shift in how work gets done.

Several major technology and commercial shifts are now converging: 

Amazon has deployed more than one million robots across its operations network and introduced DeepFleet, a generative AI foundation model designed to coordinate robot movement and improve fleet travel efficiency by 10%. 

Google DeepMind is advancing robotics foundation models and vision-language-action systems through Gemini Robotics, which it describes as enabling robots to perceive, reason, use tools and interact with humans across different robot forms. 

Nvidia and ABB are investing heavily in simulation, synthetic data and digital twins. ABB’s RobotStudio HyperReality, enabled by Nvidia Omniverse libraries, is designed to close the sim-to-real gap and support virtual training before real-world deployment. 

SoftBank is also making one of the most consequential strategic moves in the sector. It has agreed to acquire ABB’s Robotics division for $5.375 billion, subject to regulatory approvals, with completion expected in mid-to-late 2026. SoftBank frames the acquisition as part of a broader strategy across AI chips, AI robots, AI data centres and energy. 

This matters because SoftBank’s play is not just a bet on robots. It is a bet on the full Physical AI stack: compute, industrial robotics, AI models, data infrastructure, deployment channels and operating intelligence. ABB said its robotics division generated $2.3 billion in 2024 revenue and employed around 7,000 people, making this a major industrial platform rather than a speculative robotics start-up. 

The organisations that succeed in this next phase will not be defined by machine capability alone. They will be defined by how effectively they redesign workflows, integrate robotics into operations, create safe human-machine hand-offs, manage fleets reliably at scale and capture proprietary real-world data. 

From automation to physical intelligence

Not all robotics markets are maturing at the same speed. 

Industrial robotics, autonomous mobile robots and collaborative robots are already delivering measurable value in structured environments such as manufacturing, logistics and warehousing. 

Humanoids and embodied AI systems remain strategically important, but are commercially earlier-stage. 

The near-term value remains concentrated in repeatable operational environments where workflows, safety boundaries, economics and data capture are clear. 

Where value will accrue

The key question is no longer whether robotics adoption will grow. It already is. 

The more important question is where value will sit across the robotics value chain. 

Historically, competitive advantage centred on hardware engineering and automation capability. Increasingly, however, value is shifting toward the operational layer around the robot: 

• Workflow redesign  

• Systems integration  

• Fleet orchestration  

• Simulation and digital twins  

• Safety governance  

• Servicing and uptime management  

• Robotics-as-a-Service  

• Data collection, labelling and feedback loops  

• Proprietary operational intelligence  

In other words, the bottleneck is increasingly the operational system around the robot, not the robot itself. 

The most valuable robotics businesses and robotics-enabled operators are likely to be those that can combine deployment, data and continuous learning. This is especially true in service environments, where physical work is repetitive, distributed and rich in operational variation. 

The strongest business cases typically combine multiple outcomes simultaneously: labour leverage, uptime, quality, safety, compliance and proprietary learning. 

The data bottleneck

The most important strategic issue in robotics may be data. 

Simulation, synthetic data and digital twins matter. They can reduce deployment risk, accelerate testing and lower the cost of training. But simulation is unlikely to be enough on its own for robots operating in open-ended, messy, human environments. 

The physical world is difficult because it is full of edge cases: lighting variation, clutter, surface textures, force, friction, fragile objects, unexpected human behaviour, safety exceptions, awkward layouts and changing operating conditions. Robots do not simply need to “see” the world. They need to act in it, safely and repeatedly. 

That makes real-world robotics data far more valuable than generic digital data. Large language models benefited from internet-scale text and image data. Robotics does not have an equivalent internet of physical action, for example, we might contrast LLMs trained on internet-scale data with robotics foundation models trained on far smaller volumes of robot trajectories, and describing robotics as a data collection challenge.  

This means the next robotics moat may sit with whoever can collect, own and structure the best real-world operational data. 

That includes: 

• Human task demonstrations  

• Teleoperation logs  

• Robot failure modes  

• Intervention events  

• Sensor streams  

• Building maps and spatial data  

• Maintenance records  

• Safety incidents and near misses  

• Workflow exceptions  

• Task completion data  

• Environmental context  

• Human feedback and escalation decisions  

In practice, this makes operating companies strategically important. The companies closest to physical work may have the best data advantage — if they design for it. 

The Shift example: data collection becomes a business model

The emergence of companies like ‘Shift’ shows how quickly real-world robotics data is becoming a market. 

Shift is offering free home cleaning in New York in exchange for recording cleaners as they scrub, vacuum, dust and wash, using camera-equipped headwear to capture first-person task data for robotics training. 

This is a striking signal. It suggests that the market is beginning to value physical-world data highly enough to subsidise real services in exchange for it. The Verge also reports that Shift says it already pays people across 15 countries to record activities through its app, with plans to expand the model beyond cleaning into areas such as plumbing, cooking and building. 

This resembles the early gig economy in an important way. In the last decade, gig labour helped digital platforms scale by turning distributed human activity into structured supply. In robotics, gig-style human work may become a way to collect the embodied data needed to train machines. 

But the analogy also highlights the governance challenge. Physical-world data is intimate. It may include homes, workplaces, faces, screens, private spaces and sensitive routines. The companies that win will not only collect the most data. They will need to collect it with permission, structure it effectively, protect privacy and convert it into safe, reliable automation. 

SoftBank’s Physical AI play

SoftBank’s move into ABB Robotics is significant because it signals a shift from robotics as an isolated hardware market to robotics as part of a broader AI infrastructure stack. 

SoftBank says the ABB acquisition is intended to strengthen its AI robotics business and complement existing robotics-related investments including SoftBank Robotics, Berkshire Grey, AutoStore, Agile Robots and Skild AI. 

That is an unusually broad footprint. It spans industrial robotics, logistics automation, embodied AI research, robotics platforms and the compute infrastructure needed to support AI systems. 

The strategic logic is clear. If AI is moving from the digital world into the physical world, then the winning players may need more than models. They may need: 

• Robotics hardware  

• Industrial deployment channels  

• AI chips and compute  

• Data centre capacity  

• Energy access  

• Simulation environments  

• Real-world data pipelines  

• Fleet operations capability  

• Customer relationships in physical industries  

SoftBank appears to be assembling pieces of that stack. 

For operating companies, this is an important signal. The robotics market may not evolve as a simple vendor-customer relationship where operators buy machines from hardware companies. It may become an ecosystem contest between vertically integrated players, AI infrastructure companies, robotics foundation model providers, and operators that own the real-world data layer. 

Why adoption still stalls

Reliability still matters more than demos

Humanoids and embodied AI systems attract significant media attention, but scaled commercial deployment remains limited. Many general-purpose applications are years away, with nearer-term value concentrated in vertical applications that can work with today’s algorithms and data collection methods.  

This distinction matters. A compelling demo is not the same as a reliable operating model. 

Dexterity remains difficult

Progress in robotic manipulation is accelerating, but reliable operation in unstructured environments remains challenging. Google’s Gemini Robotics-ER 1.6, for example, is designed to help robots interpret visual data, perform spatial reasoning and plan actions from natural language commands, but Google also notes that physical robots can cause damage and that safety in real-world robotics remains an active area of research. 

Simulation reduces risk, but does not remove the need for real-world learning

Simulation can materially reduce cost and speed up deployment. ABB and Nvidia’s work on RobotStudio HyperReality is aimed precisely at closing the gap between virtual training and real-world deployment. 

But for many service environments, especially those involving humans, homes, workplaces or changing physical conditions, the final learning loop still has to happen in the real world. 

Integration economics can stall deployment

Many robotics projects fail not because the robot cannot perform the narrow task, but because the surrounding operational complexity becomes too costly or difficult to manage. 

System integration, workflow redesign, safety processes, workforce adoption, exception handling and ongoing support often create larger barriers than hardware acquisition itself. 

Data is expensive

Real-world robotics data is hard to collect, hard to label and hard to generalise across environments. Data is expensive and without an algorithmic breakthrough, training generally useful robots may remain prohibitively costly.  

This is why the data layer should be treated as a strategic asset from the beginning, not as an afterthought. 

The robotics operating model

Organisations should treat robotics as an operating model capability, not simply a technology investment. 

The most effective robotics strategies are likely to focus on seven areas. 

1. Start with the workflow

Focus first on high-value, repeatable operational environments where economics, ownership and safety boundaries are clear. 

2. Redesign the process

Robots should sit inside redesigned workflows with defined inputs, outputs, exceptions, escalation paths and service levels. 

3. Use simulation early, butvalidatein the real world

Simulation, digital twins and synthetic environments can reduce deployment risk and accelerate learning cycles. But organisations should not confuse simulated performance with operational readiness. 

4. Build governance up front

Safety, escalation paths, accountability, privacy, data rights and workforce transition plans should be designed into deployment from the beginning. 

5. Prioritise fleet operations

Long-term value often sits in uptime management, servicing, monitoring and continuous optimisation. Robotics at scale is closer to running an operational network than buying a piece of equipment. 

6. Build the data flywheel

Every deployment should be designed to generate useful learning data: task completions, interventions, failures, exceptions, environmental maps, maintenance records and human feedback. 

7. Rethink commercial models

Robotics-as-a-Service, outcome-based pricing and data-sharing models are likely to become increasingly important as adoption scales. In some markets, the service itself may even become a mechanism for data collection, as the Shift example suggests. 

Strategic questions leaders should ask

Leadership teams should focus on a small number of strategic questions: 

• Where are our most structured operational environments?  

• Where is labour constrained, costly or difficult to retain?  

• Which workflows create the greatest operational friction?  

• Which tasks generate repeatable, high-quality physical-world data?  

• Do we own the data generated by our operations, vendors and robotics deployments?  

• What consent, privacy and governance models are required to use that data?  

• Where can simulation reduce deployment risk, and where is real-world validation essential?  

• Where could we own part of the operational layer rather than simply purchase hardware?  

• Which robotics use cases can create a compounding data flywheel?  

• Where should we operate, orchestrate or eventually produce robotics capability ourselves?  

The organisations that answer these questions effectively are more likely to scale robotics beyond isolated pilots. 

Final takeaway

Physical robotics will not create value simply because robots are becoming more capable. 

It will create value when organisations make those systems useful, reliable, governable and economically viable inside real operational workflows. 

The next phase of competition will not be won by hardware capability alone. Nor will it be won by simulation alone. 

It will be won by organisations that can combine physical operations, safe deployment, fleet orchestration and proprietary real-world data. 

The strategic question is therefore not just: “Which robots should we buy?” 

It is: “Which parts of the physical operating system do we want to own?”