The experience, development tools and learned capabilities that let machines adapt to tasks.
AS OF 2026-09-13DATED · NON-EXHAUSTIVE
Research reviewed 2026-09-13. Available means documented access or production, not assured spare capacity for every specification. Tight refers to the scope identified in the node, not every robot component. Supply unclear means the reviewed evidence does not settle availability at the selected boundary. Expansion describes how additional usable supply can be delivered, not a promised lead time. Pricing power describes our assessment of the ability to defend price or terms; it is not inferred from technical importance alone.
Teaching machines useful behaviour
Experience has to match the work
A robot learning to grasp a deformable bag needs examples of what happened when it acted, not just pictures of bags. The useful record connects observations, commands and outcomes. Collecting that experience, making it searchable and turning it into behaviour create different purchasing decisions.
Physical Intelligence's February 2025 openpi release provides model weights and adaptation code, while explicitly warning that the model may fail on a new setup. The reusable starting point and the work needed to make it useful should remain separate in the economic analysis. A general policy is not yet a commissioned task.
This map keeps data collection and preparation together as a supplied service, and logging, search and visualization together as a development platform. It does not promote every operation in a training pipeline into a market. Simulation supplies another source of experience; training compute and deployed runtime have their own branch pages.
More useful experience, fewer repeated mistakes
Demand can grow with more robot programmes, more varied tasks within a programme, and more deployed experience worth investigating. Those mechanisms need not track robot shipments. One difficult new task can require substantial data work before a machine is sold; a mature application may reuse a policy across many installations.
Better starting models can reduce the demonstrations needed for adaptation while encouraging more teams to attempt new tasks. That creates two opposing effects for data services: fewer purchased examples per successful adaptation, but more adaptations. For tools, the corresponding question is whether useful data volume and engineering usage grow faster than prices fall.
Foxglove cut its self-service entry price in April 2026 and introduced volume discounts. It is a concrete example of a supplier encouraging adoption through lower unit prices. Growth in a robotics workflow does not, by itself, establish pricing power.
Supply-chain map
From experience to a supplied capability
Data services and development tools support policy work. Task software can use general policies or its own narrower methods. The dashed route is an option, not a compulsory model purchase.
Arrows show functional use, not contracts between the named companies. Captive data collection and in-house model development remain alternatives. Robot data platforms support both training and investigation after deployment; the map shows the learning-facing use.
Training data services
Supply unclear
Role
Supplies usable training examples
Expansion
Collection operations and acceptance checks
Pricing power
Task expertise competes with automation
Why it exists
The customer buys usable examples under a specification: the task performed, the views captured, the action and timing information, and the quality required. Collection, annotation and quality control belong together when the supplier is responsible for the accepted dataset. Capturing footage without the right measurements does not discharge that responsibility.
Relevant businesses
Scale AI (private): physical data collection, demonstration datasets and grounding annotations; its offering includes operated collection platforms.
Sama (private): robotics and manufacturing data annotation and model-validation work, including multi-sensor sequences. This is a narrower role than operating every kind of robot collection site.
Worth understanding
The distinction from a data platform is the delivered work. A customer may license tools and run its own collection team, or buy accepted examples from a service provider using its own tools. These are alternatives in the make-versus-buy decision, not two mandatory payments for the same dataset.
Scale's collection offering includes calibrated platforms and programme-specific data specifications. Expanding that service means more than adding annotators: task access, equipment, supervision and acceptance checks must expand together. Sama describes automated preprocessing alongside human review, showing how software can reduce the labour required per accepted example.
Our read is that difficult collection settings and reliable acceptance can support differentiation, while easier annotation is exposed to automation and supplier comparison. The reviewed service descriptions do not establish current spare capacity, achieved contract prices or an industry-wide queue. Supply is therefore unresolved at this boundary, rather than presumed scarce because robot data is technically valuable.
Engineers need to find the episode where a robot failed and reconstruct what its sensors and controller saw. A robot data platform supplies logging interfaces, synchronized inspection, search and data access. Those functions form one useful workflow purchase; separate boxes for logging, visualization and querying would obscure the integrated product.
Relevant businesses
Foxglove (private): robotics data collection, visualization, debugging and curation software.
Rerun (private): an open-source SDK and commercial Hub for logging, querying and managing multimodal physical data.
Worth understanding
Foxglove's April 2026 pricing change reduced Pro's starting price from $126 to $20 a month and separated storage, querying, indexing and bandwidth meters. Those are published self-service terms, not an estimate of enterprise contract prices. They show both an actual monetization mechanism and price reductions within it.
Rerun separates its SDK from the Hub backend. That distinction makes open development tools a route into a paid service without making every user a paying customer. The supplier has to save enough engineering effort or support enough scale to justify the commercial layer.
Available here refers to access to the offered software, not a guarantee that a customer's data is already compatible or that hosted service has unlimited capacity. Our inference is that integration and accumulated workflows can make replacement costly, but open interfaces, self-hosting options and falling usage prices constrain extraction. Bytes collected are a poor revenue proxy unless they also create paid processing or access.
A policy maps observations and instructions into actions. A general robot policy attempts to reuse learned behaviour across tasks or embodiments, reducing how much must be learned separately. Its supplied scope can be weights, code and adaptation tools; the customer still has to connect it to a particular machine and establish useful performance.
Relevant businesses
Physical Intelligence (private): openpi code and pretrained policies, with adaptation examples and collaboration on specific uses.
NVIDIA: GR00T robot-policy weights and post-training support for robot- and task-specific data.
Worth understanding
The openpi release distinguishes a base model, task-tuned checkpoints and further fine-tuning. Its examples establish possible reuse, not dependable performance on arbitrary machinery. The GR00T model card permits commercial and non-commercial use under its stated licence and describes post-training for specific embodiments.
Available refers to released model assets. Qualified behaviour remains a separate engineering result. Copying weights scales cheaply; improving the model or adapting it to unfamiliar tasks can consume data, compute and robot time. A single supply label cannot turn the first fact into a claim that the second is solved.
Our economic read is that broad reuse can shift value toward adaptation, distribution and complementary platforms even if the base weights earn little directly. Paid partnerships are possible, but the reviewed releases do not establish a durable general-policy licence price. This node remains useful because it changes the cost and ownership of downstream task development, not because a settled merchant market is assumed.
A builder can purchase a more specific capability, such as finding and picking an object, rather than adapting a general policy itself. The product combines the algorithms and interfaces needed for that defined job. Its boundary is narrower than a complete installed workcell and broader than an unadapted model.
Relevant businesses
Siemens: SIMATIC Robot Pick AI, a supplied picking capability with specified interfaces and licensing.
Intrinsic, a Google / Alphabet business: Flowstate and combinable skills for building robot applications, including perception, planning and sensor-based control.
Worth understanding
Siemens' Robot Pick AI listing describes a one-time, per-instance licence. Intrinsic's July 2026 technical specifications describe tools and starting components that experienced developers configure into solutions. These are different delivered scopes, even when both reduce application-development work.
The buyer compares licence and integration cost with developing the behaviour internally. Our inference is that repeatable deployment across similar tasks can improve supplier economics, while each new object set, sensor arrangement or recovery requirement can bring engineering work back into the sale. The commercially useful unit is the supported capability, not every algorithm inside it.
The reviewed product material establishes offerings, not the present capacity of implementation teams or accepted output across all applications. Nor does a licence prove pricing power: a vendor must retain value as competing software and stronger base models lower development cost. General policies feed this route only when the chosen implementation uses them.
The strongest commonality is the repeated cost of understanding machine experience and adapting behaviour. It can support a reusable tool or capability across many robot builders without making their raw data interchangeable. A supplier's useful reach is therefore wider than one family, but narrower than every instance of physical intelligence.
Our read is that the attractive position would combine repeatable delivery with customer-specific knowledge that remains difficult to replace. The principal counterforce is improvement in the tools themselves: open models, automated annotation and easier integration can expand the market while reducing spending per task. Company-level work would need to establish which effect dominates paid usage and retained margins.
This dated map selects economically useful boundaries. Company participation does not establish a trade relationship between the named suppliers, and a product announcement is not evidence of accepted machine output.