PhysDBPhysical AI Map

Trust Layer

Refusal and correction are part of the product.

PhysDB begins with trust assets because Physical AI is easy to overstate.

Anti-misinterpretation registry

vla-is-not-robot-brain

A VLA or robot foundation model can be described as a complete robot brain.

It hides the rest of the stack: embodiment, actuation, sensing, safety, latency, control, data, deployment environment, and maintenance.

Call it a model or policy component inside a robot stack, unless a source explicitly describes the shipped system boundary.

simulation-equals-real

Performance in simulation proves real-world readiness.

Simulation is useful for data, validation, and stress testing, but real friction, contact, sensor noise, wear, and human environments can break assumptions.

Simulation can reduce iteration cost and expose candidate failures; real deployment still needs measured validation.

graph-centrality-is-importance

A high-degree PhysDB node is the most important concept in Physical AI.

Graph degree reflects current corpus coverage, not intrinsic importance or causal priority.

High-degree nodes are graph leads or bridge candidates within the current source set.

demo-equals-capability

A public robot demo proves broad capability.

Demos often use curated scenes, known objects, controlled lighting, limited task distributions, or operator recovery.

Treat demos as evidence of a tested scenario, not as a general capability claim.

physical-ai-is-only-humanoids

Physical AI means humanoid robots only.

Autonomous vehicles, manipulators, drones, warehouse systems, surgical and inspection platforms, and embodied perception systems share parts of the same stack.

Humanoids are one embodiment class inside the broader Physical AI map.

Correction ledger

2026-06-09

Correction ledger opened before first public release.

PhysDB should start with correction infrastructure, not add it only after mistakes happen.

Source registry

NVIDIA Cosmos 3 technical blogNVIDIA / Cosmos 3 physical AI model framing, world generation, physical reasoning, and action generationCosmos 3: Omnimodal World Models for Physical AINVIDIA research authors / Technical report for Cosmos 3 model family and modality framingRT-2: New model translates vision and language into actionGoogle DeepMind / Vision-language-action model framing and robot-control transfer from web and robot dataRT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic ControlGoogle DeepMind research authors / VLA model architecture and problem statementOpen X-Embodiment: Robotic Learning Datasets and RT-X ModelsOpen X-Embodiment collaboration / Cross-embodiment robot datasets and RT-X model experimentsGemini Robotics brings AI into the physical worldGoogle DeepMind / Gemini Robotics and Gemini Robotics-ER model framingPI0: A Vision-Language-Action Flow Model for General Robot ControlPhysical Intelligence research authors / Generalist robot policy and VLA flow-model framingOpen Sourcing PI0Physical Intelligence / Open PI software release and fine-tuning boundary languageIsaac GR00T developer pageNVIDIA / Robot foundation models, simulation, data pipelines, and onboard compute stackLeRobot: An Open-Source Library for End-to-End Robot LearningHugging Face and collaborators / Open robot-learning stack, datasets, motor control, storage, and reproducibilityROS 2 documentationOpen Robotics / ROS community / Robotics middleware and software integration contextMuJoCo documentationMuJoCo project / Physics simulation and dynamics engine context