Founding Note: Physical AI Is A Stack, Not A Model
Published June 2026
The Gap
Robotics has entered the foundation-model era, but the public language around it is already too flat.
A robot is not a language model with arms. It is a body under force, contact, delay, heat, wear, calibration drift, sensor noise, task ambiguity, and safety constraints. A capable model can be a decisive part of the system, but it is still only one part of the system.
That is the reason PhysDB exists.
We need a map that can hold the whole stack at once: world foundation models, vision-language-action policies, robot data, simulation, digital twins, physics engines, sensor fusion, control loops, onboard compute, benchmarks, and safety evaluation.
The current source trail shows why this map now matters. Google DeepMind's RT-2 work framed a robot-control model as a vision-language-action model, connecting web-scale vision-language pretraining to robot actions. Open X-Embodiment then pushed the question across robots, assembling data from 22 different robots and 21 institutions, with 527 skills and 160,266 tasks reported in the paper. Physical Intelligence's PI0 line and NVIDIA's GR00T/Cosmos stack show the same pressure from different directions: the model is moving closer to the physical loop, but the physical loop still has many layers.
Why Now
Three changes make Physical AI a useful domain for a dedicated knowledge graph.
First, the model interface is changing. VLA models and world models are not just classifying images or writing text; they are trying to output actions, predict scenes, reason about physical state, or generate training context.
Second, robot data is becoming a first-class asset. A policy trained on a narrow robot, task, and room does not automatically become a general robot worker. Cross-embodiment datasets, data standards, open tooling, and simulation pipelines now sit near the center of the problem.
Third, evaluation is harder than demos make it look. A robot can fail because it misunderstood an instruction, because its camera missed a state change, because its gripper touched at the wrong angle, because inference arrived late, because heat throttled compute, or because the environment changed after planning.
PhysDB treats these as graph edges, not footnotes.
What PhysDB Will Study
PhysDB starts with five questions:
- Which parts of the Physical AI stack are model problems, and which are embodiment, data, control, or validation problems?
- Where do world models help most: synthetic data, physical reasoning, simulation variation, or evaluation?
- What does cross-embodiment learning actually transfer, and what stays hardware-specific?
- Which claims need a real-world measurement before they can be trusted?
- How should graph centrality be read without turning corpus coverage into a fake ranking of importance?
Operating Rule
PhysDB will keep source-backed relationships close to their evidence.
A VLA model can be a policy component. It is not a complete robot brain unless the system boundary is explicit. Simulation can accelerate iteration. It is not proof of real-world safety. A demo can reveal a tested scenario. It is not a general capability certificate.
The first version of PhysDB is small on purpose: a working graph, a source registry, an anti-misinterpretation registry, and a correction ledger opened before launch.
The question is not whether robots will become more capable. The question is which parts of the stack become reliable, measurable, and reusable enough for capability claims to survive contact with the world.