Physical Intelligence System Stack

Definition

Physical intelligence system stack is the complete set of data, compute, models, algorithms, chips, software, hardware, sensors, controls, scenes, solutions, operations, and feedback needed for an intelligent machine to perceive, decide, and act reliably in the physical world.

Current Synthesis

The bounded sources reject reducing physical intelligence to a model or robot body. Yao Song / 姚颂 emphasizes vertical system coverage from data and compute through hardware, scene access, solutions, and remote operation. 地平线 adds a time-coupling constraint: algorithms can change quickly, while chips take years to define and must operate within edge power, heat, space, and latency limits. Liber AI adds a competitive boundary: general foundation models may absorb more semantic understanding and planning, while action generation, contact feedback, motion control, embodiment, hardware reliability, and scene delivery remain specialized. System performance therefore depends on interfaces and co-design across layers, not simply owning every layer.

Cars and robots instantiate the stack differently. Vehicles offer scaled production and tight real-time safety requirements; robots add varied bodies, joint control, contact, manipulation, and task-dependent thinking time. A platform can reuse design and data-loop methods across them, but must still adapt hardware, control, and evidence to each embodiment. Vertical integration is therefore selective rather than absolute: control the layers whose interfaces materially affect learning or reliability, and use mature suppliers where ownership would add cost without improving the system.

Key Claims

• Space perception, body motion, manipulation, and remote operation are all parts of physical intelligence rather than separate afterthoughts.
• Vision Language Action Models, World Models, World Action Models, and Robot Reinforcement Learning are route components, not complete company definitions by themselves.
• Scenario access matters because physical-world data and commercial value appear only when robots act in real settings.
• Early vertical integration can be rational when standard interfaces, mature suppliers, and stable form factors have not yet formed.
• The stack view supports Milestone Commercialization: each technical step should be linked to a deployable service or customer value where possible.
• Chip roadmaps and model roadmaps must be coordinated because silicon decisions precede the software workloads they will eventually run.
• Edge power, heat, space, latency, safety, and embodiment constraints determine which nominal capability becomes usable action, so general models can move deeper into planning without eliminating specialized policies, low-level control, sensors, hardware adaptation, or customer delivery.

Evidence

Full-stack scope and scene access

• 173: 对话姚颂：深鉴、东方空间、再出发，「天才少年」十年后 defines physical intelligence through data, compute, models, algorithms, software, hardware, scenarios, solutions, and remote systems, and connects scenario access to both learning and commercial value.

Hardware–software time coupling

• 特番｜从会开车到会做事：AI如何进入真实世界 says vehicle chips may take three to four years from definition to deployment, so algorithm teams must anticipate future model characteristics while chip constraints bound later software capability.

Embodiment-specific system design

• 特番｜从会开车到会做事：AI如何进入真实世界 contrasts edge vehicle requirements with data-center compute and distinguishes robot high-level decisions from joint-level control, balance, tactile feedback, and body limits.

Layer boundary and selective integration

• 187: 对话LiberAI刘松铭：Astra来袭，00后物理AI创始人看到的胜率 divides the robot stack into System 2 planning, System 1 action, and System 0 motion control, then argues for owning model-relevant data, hardware, control, and deployment links while sourcing mature components elsewhere.

Counterevidence & Qualifications

• Full-stack thinking does not prove that one firm should vertically own every component; modular suppliers and standards can outperform integration when interfaces mature.
• The sources are founder or company accounts and do not independently compare stack ownership, platform economics, or performance.
• Co-design can improve fit while increasing capital intensity, roadmap lock-in, organizational complexity, and exposure to wrong long-range forecasts.
• The System 2/System 1/System 0 boundary is a useful engineering and company heuristic, not a stable industry standard; general models, specialized policies, and control systems may change their division of labor.
• Liber AI’s reported model, hardware, and deployment advantages are founder claims and do not establish that vertical integration will outperform modular competitors at scale.

What Changed

• Added the System 2/System 1/System 0 division and the competitive boundary between general models and physical execution.
• Clarified selective vertical integration: own interfaces that determine learning and reliability, not necessarily every component.
• Added customer delivery and scene infrastructure as durable stack layers alongside models and hardware.

Related Concepts

• Physical AI - broader physical-world intelligence category instantiated by the system stack.
• Embodied AI Value Chain - distribution of value across the stack’s commercial layers.
• Physical World Data Flywheel - feedback mechanism connecting deployed action to later improvement.
• Production Robot Scenario Selection - scene choice that determines useful data and customer value.
• Hardware–Software Iteration Asymmetry / 软硬件迭代不对称 - timing and reversibility difference between physical and digital layers.
• Embodied Intelligence Vehicle Path - vehicle-first route for scaling and financing the stack.
• Robot Form-Factor Pragmatism - body-selection rule that can reduce unnecessary stack complexity.
• General Model Robot Boundary - changing division of labor between broad foundation models and embodied specialization.
• Layered Robot Architecture - technical separation of slower planning from low-latency physical execution.