Methodology
How the Atlas is built
The Physical AI Atlas is a sourced company-level dataset, not an audited financial database or an estimate of total market revenue. Its purpose is to make evidence, uncertainty and differences between products visible.
Updated 30 September 2026
Company inclusion
A company is included when it builds a physical autonomous system or a material enabling layer for Physical AI and has enough public evidence to support a useful profile.
The Atlas covers robot makers, autonomy and embodied-model companies, and edge-compute suppliers. Categories describe the company’s primary role in the Atlas; they do not assert that a company operates in only one market.
A company is not added solely because it invests in robotics, partners with a robot maker or supports a compatible software framework.
Product maturity
Product status is based on public evidence for real products and deployments, not the maturity of a company’s marketing narrative.
Shipping means a product is buyable or in series production. Deployed pilots means units are operating in real environments without establishing broad availability. Announced and in development remain forward-looking.
Status is not inherited between models. A shipping earlier generation does not make a newly announced generation shipping.
Software and compute hardware
The data model keeps the robot’s intelligence stack separate from the silicon or computer used to execute workloads.
AI and software stack records models, policies, runtimes, operating systems, simulation and cloud services.
Compute hardware is product- or configuration-specific. Exact chip models are stored only when a source states the exact SKU. Training GPUs, external computers and onboard processors receive different roles.
Use of NVIDIA Isaac, an investment by a semiconductor company or an Arm CPU architecture does not independently establish the onboard silicon vendor.
Funding and valuation
A number is not comparable until its meaning and date are known.
The Atlas distinguishes single rounds, cumulative equity, committed capital, debt, grants, IPO proceeds and conditional subscription amounts. Amounts with “more than” or “up to” retain those qualifiers.
Round values are not added mechanically. Public-company market capitalization and parent-company assets are not mixed into startup-style funding totals. A value without a recorded field-level source is labelled “Source not recorded”.
Events and updates
One announcement is one event, even when it names several companies.
The event date records when the development occurred; the checked date records when it was verified. Funding, contracts, partnerships, production, deployments and model releases remain separate event types.
Partners and investors are not promoted to customers. Planned production volumes are not counted as installed units, and contract ceilings are not treated as funding.
Geography, sourcing and limits
The Atlas compares Europe, the United States and China as geographic poles. Europe is not equivalent to EU-27 and includes non-EU European headquarters.
Sources are marked as company-stated, independent or estimate. Company statements can establish what a company announced; they do not provide independent validation of performance.
A missing field means that no acceptable evidence has been recorded. It does not prove that the fact is false. Coverage varies by category and should be reported alongside any aggregate derived from the dataset.
Use the data responsibly
Every aggregate inherits the limitations of the underlying records. Open company profiles for field-level dates, sources and caveats before using a figure in investment, procurement or policy work.
The public JSON endpoint is available at /data/companies. Reuse should preserve attribution, dates, qualifications and exclusions.