Physical AI: State of Play
From foundation models to factory floors, a map of the robotics and embodied AI landscape and the metrics that will decide it.
Executive summary
- Robot foundation models are replacing task-specific programming and making general-purpose machines plausible.
- Deployment is concentrated in structured industrial and logistics settings; humanoids are in pilots.
- Cost per task and supervision ratio are the decisive economics, driven by fleet learning.
- The field is crowded and well funded; deployment data will drive consolidation.
- Safety certification is the regulatory frontier.
Key findings
- 01Fleet learning improves supervision ratios faster than hardware improvements reduce cost.
- 02Simulation is necessary but real-world data is the moat.
- 03Operators are shifting evaluation from capability demos to task-level metrics.
- 04Humanoids win where flexibility beats throughput; wheeled and fixed platforms win elsewhere.
Physical AI is the broadest and slowest of the trends on the Parallax Index, and the one with the largest eventual footprint. This Deep Dive maps the stack from models to machines, distinguishes what is deployed from what is demonstrated, and sets out the metrics we will use to judge progress.
1. The stack
At the bottom sit sensors, actuators and onboard compute. Above them, control software that keeps a machine balanced and moving. Above that, perception and vision-language-action models that translate what the robot sees and is told into actions. Around all of it: simulation environments for training, fleet software for learning across machines, and safety layers that bound behaviour.
2. What is deployed versus demonstrated
| Application | Status | What to watch |
|---|---|---|
| Warehouse picking and material handling | Deployed at scale (mobile robots); humanoid pilots | Supervision ratio, cost per pick |
| Machine tending | Pilots | Uptime, changeover time |
| Assembly | Early pilots | Precision, cycle time |
| Inspection | Deployed (drones, mobile) | Coverage, detection rate |
| Hospital logistics | Pilots | Safety incidents, staff acceptance |
| Home tasks | Research | Reliability in unstructured settings |
3. The economics
The decisive number is cost per task, fully loaded: hardware amortisation, energy, maintenance and the human supervision time per machine. The supervision ratio, how many robots one person can oversee, drives that number more than hardware price. Fleets that learn from each other improve the ratio over time; isolated deployments do not.
Illustrative cost per task versus supervision ratio. Cost per task index: 1:1 100, 1:3 62, 1:5 48, 1:10 36, 1:20 30.
4. The players
Chipmakers and simulation providers supply the platform. Humanoid startups and established robotics firms build machines. Automakers and logistics operators are both customers and, in some cases, developers. Frontier labs contribute models and, increasingly, partnerships. The field is crowded and well funded; consolidation is likely once deployment data separates contenders.
5. Risks
- Safety in shared human environments and the certification frameworks that govern it.
- Hype outrunning deployment data, leading to capital misallocation.
- Hardware reliability and the cost of maintenance at fleet scale.
- Supply chain dependence for actuators, sensors and batteries.
- Labour and political response to visible automation.
6. Opportunities
- Fleet software and data platforms.
- Simulation and synthetic data.
- Component supply chains for actuators and sensors.
- Integration and operations services for deployers.
- Safety certification and testing.
7. What happens next
Watch for published cost-per-task and uptime data from multi-site deployments; safety certification frameworks for robots in shared spaces; and consolidation among humanoid contenders. The industrial and logistics use cases will scale first. Humanoids will follow where dexterity and mobility justify the cost.
- 2010sDeep learning perception
Robots learn to see reliably.
- 2022Language grounding
Models connect instructions to actions.
- 2024Robot foundation models
Generalist policies across robots and tasks.
- 2025Humanoid pilots
Paid deployments in logistics and automotive.
- 2026Commercial metrics
Cost per task and uptime enter the conversation.
Market map
Technology explanation
Sensors feed perception models; vision-language-action models map perception and instructions to motor commands; control software executes them safely; simulation and fleet learning supply training experience at scale.
Risks
- Safety in shared spaces
- Hype versus deployment data
- Fleet-scale reliability
- Supply chain dependence
- Labour and political response
Opportunities
- Fleet software
- Simulation and synthetic data
- Component supply chains
- Integration services
- Safety certification
What happens next?
- Published deployment metrics
- Safety frameworks
- Consolidation
- Industrial scaling first
Related topics
Sources & references
- 01Robot learning and vision-language-action literature — Academic and industry labs; see Sources pageresearch
- 02Deployment announcements from robotics vendors and operators — Company press releasescompany
More from Artificial Intelligence
The Agentic Enterprise: From Pilots to Production
A Deep Dive on enterprise agent deployment: the adoption curve, the market map, the technology, the risks and the opportunities.
Open-Weight Models Are Becoming Strategic Infrastructure
Open-weight releases shape who can build, where models run and which countries control their AI stack. Governments and enterprises are treating them like infrastructure.
Small Reasoning Models Are the Quiet Revolution
An explainer on test-time compute, distillation and why small reasoning models make private, on-device and regulated AI deployments viable.