Why reusable physical capabilities may become the layer that lets AI agents operate many kinds of robots without controlling every joint directly.
By FlyPig AI InsightsPublished Updated Research series
Short answer
The closest Physical AI equivalent to an app may be a trusted robot skill: a bounded capability such as inspect, grasp, dock or deliver with explicit inputs, outputs, operating constraints and failure states. Agents can compose those skills into workflows while the skill layer absorbs robot-specific implementation details.
An app is only a useful analogy up to a point
A phone app usually owns a screen and a user session. A robot capability acts in a shared physical environment, may move mass, interact with people and must obey hardware and safety constraints. The more useful software unit is therefore not a screen-based app but a skill with a defined contract.
Skills turn embodiment into callable tools
A higher-level agent should not need to command joints or servo loops. It should reason in terms of capabilities exposed by the platform.
navigate(location)
inspect(asset)
pick(object)
place(object, location)
dock(charger)
handover(object, person)
report(exception)
A real skill requires trust metadata
Physical skills need declared hardware compatibility, confidence, failure modes, required sensing, operating envelope and safety conditions. A marketplace that distributes code without validating these boundaries would be far more dangerous than a conventional app store. Simulation evidence, runtime permissions and certification may become part of the platform itself.
This is where software reuse begins to compound
A reliable inspection or manipulation skill can potentially be reused across customers and compatible robot bodies. Reuse turns one-off integration work into software IP. The more that domain knowledge, telemetry and recovery logic are captured in the skill layer, the more difficult that layer becomes to commoditize.
The application economy begins when skills become portable
The decisive moment for Physical AI may not be when humanoids look impressive. It may be when a developer can build a useful physical workflow once and deploy it across multiple compatible machines. That is the point at which robotics starts behaving less like custom automation and more like a software platform economy.
Evidence review · September 12, 2026
The strongest counter-thesis is now end-to-end in-context learning
Confirmed evidence. Ambi Robotics has made a production-derived AI Skill Suite available for licensing through AmbiOS and says the software isolates hardware complexity across third-party robot configurations. At the same time, Skild's S1 foundation model can take a single video demonstration as context and execute unseen, long-horizon tasks without task-specific weight updates. On September 9, Skild also reported more than 60 paying customers across multiple physical applications; those commercial figures are company-reported rather than independently audited.
FlyPig interpretation. The emerging application layer may not resolve into one software primitive. Explicit skills remain attractive for permissions, testing, safety boundaries and commercial packaging, while foundation models may generate or compose behavior internally without exposing every capability as a separately installed skill. A plausible architecture is therefore a stable external capability contract wrapped around increasingly general internal models.
FlyPig prediction. FlyPig prediction FP-PAI-004 remains active at 0.60 confidence. AmbiOS and the Qualcomm-NEURA build-once model support reusable skills, while Skild S1 directly activates the prediction's end-to-end-model falsifier. Keeping confidence unchanged is more defensible than treating either architecture as the winner today.
Continue the series
Where Value Moves When Physical AI Becomes Modular
