Physical AI is scaling effort,
not knowledge

Physical AI is having its moment. Foundation models can reason about objects, plan multi-step tasks, and generate plausible manipulation trajectories from language alone. Humanoid demos ship weekly. Funding rounds break records. The intelligence is arriving.

And yet, most robotics teams still start every project from zero. New environment, new robot, new task; capture the data again, build the simulation again, train from scratch again. The work ships, but nothing carries forward. The next project begins exactly where the last one did.

This is the pattern nobody talks about: Physical AI is scaling effort, not knowledge.

Why language AI scaled and Physical AI has not

Language AI had a structural advantage that is easy to overlook in hindsight. Text is its own training data. Every book, every webpage, every conversation ever transcribed became fuel for the next model. The infrastructure to store, clean, and serve that data existed before the models did. When transformers arrived, they walked into a world already organized for them.

Physical AI has no equivalent. A robotics team that spends six months capturing kitchen environments, calibrating physics parameters, and training a manipulation policy produces something valuable; but that value is locked inside one project, one robot, one lab. The next team working on a similar problem starts from the same blank slate.

Data collection companies have made raw data more accessible; point clouds, RGB-D captures, and demonstration datasets are more available than ever. But raw data is not knowledge. Knowledge is data that has been validated against reality, structured for reuse, and made legible to systems that were not involved in its creation. That layer does not exist for Physical AI.

The failure mode has changed

Two years ago, robot manipulation failed at understanding. A robot picked up a wooden block when asked for a banana because the perception stack had no concept of what a banana was.

That problem is largely solved. Today the failure mode is grounding. The model understands “open the drawer.” It can plan the motion. But the world it trains against uses generic physics parameters; estimated friction, assumed mass, approximate joint constraints. The model executes a plan that would work perfectly on an idealized drawer and fails on the real one because the real one has wooden runners with 3x the friction the simulation assumed.

The industry workaround is domain randomization: randomize the physics parameters across a wide range and hope the real values fall somewhere inside. It works, but it is expensive, imprecise, and fundamentally wasteful. Every team that randomizes drawer friction between 0.2 and 0.8 is admitting the same thing: nobody actually measured it.

The intelligence moved up the stack faster than the world knowledge did. The models are ready. The memory underneath them is not.

What a memory layer actually means

We use the phrase carefully. A memory layer is persistent, structured, validated, and legible to other systems. It is not:

• a CAD repository (those exist, and they describe geometry without physics)
• a scene graph (a scene graph is a moment in time; memory persists across moments)
• a raw dataset (datasets record what happened; memory records what was learned)

It is:

• a structured record of how real-world objects behave; their physics, their interaction constraints, their manipulation affordances
• a trace of execution attempts; which robot, which end-effector, what the physics actually did versus what was expected
• a transfer layer that determines which robot can perform a given task on a given object, and why

The unit of this memory is what we call a neural object: a self-contained, typed bundle that carries everything a robot stack needs to interact with a real-world object without starting from scratch. The schema walkthrough covers the technical format in detail.

What this changes

When a memory layer exists, the economics of Physical AI change fundamentally. A task taught to one robot on one object in one environment becomes reusable across robots, environments, and deployments. The second team to encounter a drawer does not need to re-measure its friction, re-calibrate its joints, or re-discover its failure modes. That knowledge already exists, validated and structured, ready to be loaded.

This is what compounding looks like for Physical AI. Not bigger models or more data, but a growing body of validated world knowledge that makes every subsequent deployment faster, more reliable, and less expensive than the last. The hundredth deployment should be trivial compared to the first.

The parallel to language AI is direct: language models became powerful when they could draw on a structured, validated, ever-growing body of text. Physical AI will scale the same way; not through text, but through a structured, validated, ever-growing memory of how the real world actually works.

What we have shipped against this thesis

We are deliberate about what we claim and what we do not. Here is what exists today:

• Cabinet drawer. A trained policy opens it through real friction grip with consistent success on the calibrated asset. A separately captured human demonstration validates against the same physics. The transfer layer differentiates which robots can perform the task, with specific mechanical reasons for each verdict.
• Refillable kitchen lighter. Full pipeline from reconstruction through physics authoring to real-contact demonstration; including the articulated trigger and safety lock.
• Game controller. Articulated-row challenge: face buttons, shoulder triggers, analog sticks. Built through the family-builder path.
• Qualification suite. Deterministic, runs in CI, covers geometry through physics through integration. Every neural object must pass before it enters the memory layer.
• Transfer solver. Automated per-robot verdict with a legible failure reason, not a confidence number.

Each of these is a worked example. None of them is the product. The product is the memory layer itself and the process by which it grows.

What we are explicitly not selling

• Zero-shot transfer guarantees. The transfer layer is a hypothesis with evidence. It is sometimes wrong. It is wrong legibly.
• Universal object coverage. We handle validated object families. We do not claim to handle every object class, and we publish the family list.
• A replacement for domain randomization. Validated physics narrows the randomization range; it does not eliminate it. Edge cases (a wet floor, unexpected lighting, a worn-out joint) still require randomized training.
• A sim-to-real silver bullet. Real-grip manipulation with calibrated physics works on the assets we have shipped. It will need tuning on yours.

The fastest way to lose credibility in robotics is to oversell sim-to-real. We would rather undersell it.

The question we are working on

If Physical AI is going to reshape industries the way software did, it will not happen through brute force. It will happen when the knowledge from one deployment carries forward to the next, when validated world knowledge compounds instead of being discarded.

That requires a memory layer: persistent, structured, validated, and legible to any robot stack. We are building it. The technical deep dives on Realm and CLAP walk through the architecture.

If you are building Physical AI and this problem is familiar, we would like to hear from you.

If you want to see what this looks like in practice: What a neural object actually contains. A walk through the typed schema, the validation gauntlet, and what round-trip testing rules out.