# Glossary The terms DEPLOY uses with strict definitions. Where industry usage is loose or inflated, the definitions below reflect the registry's bar for what counts. This page is the canonical reference for the vocabulary used across DEPLOY's verified record. Numbers cited on category pages reflect these definitions. ## Brain company vs. hardware company The registry classifies companies by what they primarily build: hardware (the robot itself), brain (the AI foundation models, controllers, and software that animate robots), or both. Hardware-only examples: Boston Dynamics historically, Geely Farizon. Brain-only examples: Google DeepMind, Physical Intelligence. Both examples: Figure AI, Tesla, Rivian's Mind Robotics, all of which build both their own hardware and their own AI stack. The split matters because the physical-AI value chain is increasingly two distinct layers, and the strategic question of whether a company controls its own brain or licenses one (Apptronik using Google DeepMind's Gemini Robotics, Boston Dynamics' new Atlas using the same) is a meaningful axis for evaluating any robotics company. ## Deployment A deployment is a robot doing real economic work at a named operator site, with a published source confirming the operation. It is not a shipment, a sale, a press demo, a manufacturing run inside the maker's own facility, an announced future plan, or a memorandum of understanding. The registry's verified deployment count is restricted to operations that meet this bar at the time of listing. The distinction matters because public counts of "deployed" robots routinely include all of the above. A figure of 20,000 humanoid robots "deployed" typically reflects production volume or shipment volume across multiple categories, not robots performing economic work at customer sites. DEPLOY's smaller, stricter number is the artifact of the verification bar, not an undercount. _See also: the [methodology page](/methodology.md), [verified deployments](/verified-deployments.md)._ ## First-pass yield First-pass yield is the share of units that come off a manufacturing line meeting full specification without rework. A first-pass yield of 80 percent means 20 percent of units require teardown, repair, or reassembly before they can ship. The metric is a leading indicator of whether a robot program has crossed from prototype-with-rework into a real manufacturing operation. The threshold matters because below roughly 80 percent, the bill-of-materials cost analysis a company quotes is not the actual delivered cost: rework consumes labor, parts, and time that the published BOM does not capture. Above 80 percent, the published cost begins to approximate reality. ## Fleet A fleet is the count of active vehicles or robots a single operator currently has performing work. It is not the same as installed base, manufacturing run, shipment volume, or contracted future units. A 1,700-vehicle robotaxi fleet means 1,700 vehicles are currently operating, not 1,700 ordered, planned, or built to date. This matters because fleet figures across the industry are often quoted in ways that conflate these. The registry's fleet figures reflect current operating counts where possible, with disclosed targets recorded separately as claims. ## Form factor The form factor of a robot describes its physical configuration: bipedal humanoid, wheeled general-purpose, quadruped, industrial arm, autonomous vehicle, or fixed-pedestal. Form factor determines what work the robot can plausibly do and which category pages it surfaces under. The registry's verified-humanoid-companies category, for example, includes only models with form factor humanoid; a wheeled general-purpose robot is tracked but does not surface there. Form factor is recorded on the model record, not the company record, since a single company may build multiple form factors. ## Foundation model for robotics A foundation model for robotics is a large multimodal AI model designed to generate robot behavior across multiple hardware platforms, typically trained on combinations of human video, simulation, and real-world robot data. The category is often abbreviated VLA (vision-language-action) or robot-FM. Examples include Google DeepMind's Gemini Robotics, Physical Intelligence's pi-series models, and Skild AI's brain models. These models are the brain layer described above. A robot company that uses a third-party foundation model rather than building its own has chosen to specialize on hardware; one that builds its own foundation model is investing in vertical integration of the full stack. _See also: [Google DeepMind](/companies/google-deepmind.md), [Physical Intelligence](/companies/physical-intelligence.md)._ ## Geofenced deployment A geofenced deployment operates only inside a defined geographic boundary that has been pre-mapped and validated. The robot or vehicle cannot operate outside this boundary. Most current robotaxi services are geofenced: Waymo runs inside defined service areas in each operating city; early robotaxi services typically expand by adding new geofences rather than removing the boundary entirely. The distinction from open-world or full-autonomy operation matters when comparing operators. Two services with similar fleet sizes may have radically different geofence footprints, and the geofence determines the addressable rider market more than the fleet size does. ## Humanoid robot A humanoid robot is a general-purpose robot with a bipedal, human-scale form factor, designed to operate in environments built for humans without modification. The defining property is generality of form, not generality of intelligence. An industrial robot arm with a single anthropomorphic feature is not a humanoid in DEPLOY's classification; a bipedal robot like Apollo, Figure 03, Atlas, or Digit is. Wheeled general-purpose robots are tracked separately; the form factor field on each model record carries the precise classification. _See also: [verified humanoid companies](/verified-humanoid-companies.md), [Figure vs Apptronik](/figure-vs-apptronik.md), [Figure vs Tesla Optimus](/figure-vs-tesla-optimus.md)._ ## Incident An incident is a verified event in which a deployed robot or autonomous vehicle caused harm, behaved abnormally in a way that public sources reported, or was the subject of a regulatory action. Injury or fatality incidents require two or more independent sources for inclusion. Software bugs caught in testing, near-misses without external reporting, and unverified user complaints are not incidents in the registry's record. _See also: [verified incidents](/verified-incidents.md), the [methodology page](/methodology.md)._ ## Maturity stages DEPLOY classifies robot models by a strict five-step ladder: research, prototype, pilot, commercial, and production. Each stage is a higher trust signal than the one before it. Lab-only sales count as research, not commercial. Research: the robot exists as a study or development platform, including lab-only sales to universities and research institutions. Prototype: demonstrations, early hardware iterations, not yet trialed at a customer site. Pilot: limited real-world trials at a customer site, usually under structured supervision. Commercial: the robot is doing real economic work for a customer who is paying for the outcome. Production: mass-produced at scale, with the operating-fleet count justifying dedicated manufacturing capacity. These distinctions exist because industry usage routinely conflates them. A robot trialed at one customer site for two weeks is not "in commercial deployment." A unit produced in the maker's own factory for testing is not "shipped to a customer." The maturity field on each model carries the operative classification. _See also: the [methodology page](/methodology.md)._ ## Operator An operator is the entity that actually runs a robot or vehicle in the field, taking the commercial and operational responsibility for it. The operator may or may not be the same entity as the manufacturer. Waymo operates Waymo's vehicles; Tesla operates Tesla's robotaxis; but GXO operates Apptronik's Apollo robots in GXO warehouses, where Apptronik is the manufacturer and GXO is the operator. The registry tracks operator and manufacturer separately on each deployment for this reason. Coverage analyses, category pages, and incident attributions distinguish accordingly. ## Permit A permit, in the robotaxi context, is regulatory authorization to operate a specific type of service in a specific jurisdiction. Testing permits authorize on-road testing with a safety driver. Driverless testing permits authorize on-road testing with no driver but no passengers. Commercial passenger permits authorize paying passengers in driverless vehicles. The bar rises at each stage. A company holding a commercial passenger permit in eight countries (as WeRide does) is in a structurally different regulatory position than one holding only testing permits. The registry tracks permit type per deployment where the public record makes the distinction explicit. _See also: [Pony AI vs WeRide](/pony-ai-vs-weride.md)._ ## Physical AI Physical AI is the category of AI systems designed to perceive, reason about, and act in the physical world rather than purely digital environments. The defining property is embodied interaction: a physical AI system has sensors that read the real world, actuators or vehicles that change it, and a decision layer that connects them. Robotaxis, humanoid robots, autonomous delivery vehicles, and embodied research platforms all fall under physical AI. Enterprise language models, coding agents, and chat assistants do not. The distinction matters because investment, regulatory framing, and capability comparisons routinely conflate the two. A company building an enterprise LLM and a company building a humanoid robot are both labeled "AI companies" in casual coverage, but they face different scaling laws, regulatory risks, and customer types. DEPLOY's registry covers the physical AI category specifically. _See also: the [methodology page](/methodology.md), [verified deployments](/verified-deployments.md)._ ## Robotaxi A robotaxi is a fully driverless commercial passenger service: a passenger pays for a ride and the vehicle operates with no human safety driver on board. It is distinct from supervised driver-assistance services where a human remains in the loop, regardless of the marketing label. Tesla's Bay Area FSD-supervised ride-hailing service, for example, is not a robotaxi by this definition; Tesla's unsupervised Robotaxi pilot in Austin is. The verified record counts each robotaxi service by its current operating profile, not by the company's marketing classification. _See also: [verified robotaxis](/verified-robotaxis.md), [Waymo vs Tesla](/waymo-vs-tesla.md)._ ## Sim-to-real transfer Sim-to-real transfer is the practice of training a robot control policy in simulation and then deploying it on physical hardware, with the goal of having the simulated learning carry over to real-world performance without significant additional training. The strict version of the claim is zero-shot transfer: the policy that runs on the real robot is identical to the one trained in simulation, with no real-world fine-tuning. Looser versions involve simulation-pretraining followed by real-world supervised fine-tuning, which is a meaningfully easier task. The distinction matters because the technique is increasingly invoked as a milestone, but the strict and loose versions reflect very different engineering claims. A press release that mentions sim-to-real without specifying whether real-world fine-tuning was used should be read as the loose version by default. ## Unit economics Unit economics describes the revenue and cost profile of a single unit of operation (one vehicle, one robot) at the margin. In a robotaxi context, unit economics breakeven means a single vehicle is generating enough revenue per day to cover its allocated costs (depreciation, energy, remote assistance, maintenance, insurance), excluding company-level overhead and R&D. It is a leading indicator of when scaling the fleet generates positive cash flow rather than consuming it. A company that reports unit economics breakeven in specific cities is reporting a milestone, not company-wide profitability. The registry records these claims as claims, not as verified company-level financial facts. ## Verified A claim about a robot, company, deployment, or incident is verified when DEPLOY has confirmed it against primary sources at the bar described in the registry's methodology. Verified entries are marked with the verified badge across DEPLOY's category pages. Claims attributed to a single company press release without independent confirmation are recorded as (claimed), not verified. The split is intentional. The registry tracks both verified facts and credible claims, but distinguishes them visibly so readers and machine consumers know which is which. _See also: the [methodology page](/methodology.md), [verified humanoid companies](/verified-humanoid-companies.md) where the verified vs. claimed split is visible in the table headers._ _Last updated: 2026-05-27 ยท canonical URL: /glossary_