NovaCrest research
Where physical AI is going, and where the data has to go
Robots learn from experience of real places. To see where that experience is missing, we mapped every physical task in 923 U.S. occupations against the robot systems cited for it, and looked at where each robot has been rated able to do the work.
- of physical tasks have a robot rated able to do them somewhere
- 68%
- in everyday workplaces built for people
- 16%
- anywhere, as found, such as homes, farms and roads
- 1.4%
- with no qualifying robot yet
- 32%
Every square is a real job task
Each square is one physical task from the U.S. occupation catalogue, grouped by sector. Its colour shows the most demanding setting where a cited robot has been rated able to do it.
- Anywhere, as foundPlaces taken as they are: roads, farms, building sites and homes.
- Everyday workplacesPredictable places built for people, such as hospital corridors, warehouse aisles and kitchens.
- Built for robotsOnly where the space is built around the machine: cells, grids, conveyors and fenced floors.
- No qualifying robotNo cited robot qualifies, even in a space built around it.
Where the data needs to go
Most physical tasks have a robot rated able to do them only in robot-built settings. Bringing that work into everyday places, and reaching the tasks no robot does yet, takes experience recorded in those places.
By type of work
Share of physical tasks in each theme by the most demanding setting a cited robot reached. A task can carry several themes.
By sector, in working time
Average split of each occupation’s working time, from estimated time shares per task.
Every occupation, one map
Each dot is one of 742 occupations with physical tasks, grouped by sector. Colour it by where data is needed, search for any occupation, and open it to see how its working time splits.
Positions group similar occupations and carry no units.
What comes next
Robots can reach more work in two ways: workplaces adapt to them, or they learn work no robot does yet. Each step shows how much working time comes within reach, on the same data.
Not a forecast. These steps carry no dates, adoption rates or job effects. Work in reach is different from work replaced: cost, safety, regulation and people’s choices decide adoption.
The evidence is new
64% of dated robot citations were published in 2024 or later. Most of what is known about robots doing real job tasks is less than three years old.
Robot citations by publication year and maturity. 2026 is a partial year.
The machines behind the evidence
Every rating rests on cited robot systems. A small group of platforms appears across the whole economy, and most evidence is still indirect.
- developers named in the citations
- ≈6,000
- of citations describe systems already in real operations
- 38%
- of citations rely on a related task rather than direct evidence
- 63%
Developers cited across the most occupations
Names are grouped automatically from cited sources, so counts are approximate.
Method and sources
How the figures on this page are built.
Where the ratings come from
The task statements come from O*NET 29.3. The setting ratings, time shares and robot citations come from Anthropic’s Economic Index robot exposure release, published on 30 September 2026. They are published estimates rather than measured trials. NovaCrest regrouped and visualised them using a snapshot taken on 3 October 2026.
Working time
Each task’s estimated share of working time is split across the ways the task is done, in proportion to their weight. A part counts as rated where the work happens when the robot’s rating reaches the setting where people do the work, as robot-built only when the robot qualifies only in a more controlled setting, and as no qualifying robot when none qualifies. Sector figures are averages across occupations, unweighted by employment.
What comes next
Step 1 counts work rated where it happens. Step 2 adds work rated only in robot-built settings. Step 3 adds the remaining physical work. The steps carry no dates, costs or employment effects.
Citations and developers
Years come from each citation’s date. Developer and system names are grouped automatically from free text, so counts are approximate. Themes come from keyword rules applied to each task.
Sources for the robotics story
The milestones behind the six stages on our homepage. We chose them because they mark turning points for our work; the selection is editorial.
| Year | Milestone | What it showed | Source |
|---|---|---|---|
| Programmed and scaled | |||
| 1961 | Unimate starts work at General Motors, Ewing Township, New Jersey | The first industrial robot replayed taught positions to handle hot die-cast parts. | IEEE Robots Guide |
| 1973 | KUKA Famulus | KUKA describes it as the first industrial robot with six electric motor-driven axes, the layout most factory arms still use. | KUKA |
| 2024 | About 542,000 industrial robots installed worldwide | The International Federation of Robotics counted 542,000 installations in 2024. China accounted for 54%, about 295,000 units. | IFR, IFR China (PDF) |
| Mobile | |||
| 2012 | Amazon agrees to acquire Kiva Systems | About $775 million in cash for drive units that carry shelves to warehouse workers. | Amazon |
| 2025 | Amazon’s one millionth robot | Announced alongside DeepFleet, a model that coordinates how the robot fleet moves. | About Amazon |
| Learning | |||
| 2016 | Large-scale grasp learning | More than 800,000 grasp attempts over two months, using between 6 and 14 robot arms at a time. | arXiv |
| 2018 | QT-Opt | Over 580,000 real-world grasp attempts and 96% grasp success on objects the robot had not seen. | arXiv |
| 2019 | Rubik’s Cube with a robot hand | OpenAI trained models only in simulation, with automatic domain randomisation, and ran them on a physical hand. | arXiv |
| 2021 | Ego4D | 3,670 hours of first-person video from 931 camera wearers in 74 locations across 9 countries, built for research on egocentric perception. | arXiv |
| Generalising | |||
| 2022 | RT-1 | 130,000 episodes covering more than 700 tasks, collected with 13 robots over 17 months. | Google Research |
| 2023 | RT-2 | A vision-language-action model that learns from both web and robot data. | Google DeepMind |
| 2023 | Open X-Embodiment | Data from 22 robot types in one shared collection, which the project reports as more than a million real robot trajectories. | arXiv, project site |
| 2024 | Universal Manipulation Interface | Hand-held grippers record demonstrations in everyday places, without a robot on site during capture. | arXiv |
| 2024 | EgoMimic | Egocentric human video, paired with 3D hand tracking, used to train robot manipulation policies. | arXiv |
| Experience, now | |||
| 2025 | GEN-0 from Generalist | The company reports training on 270,000 hours of real-world manipulation data and steady gains from more data and compute in its own experiments. | Generalist |
| 2026 | NVIDIA Physical AI Data Factory Blueprint | A reference workflow for curating, generating, augmenting and evaluating physical AI training data. | NVIDIA |
| 2026 | Gemini Robotics 2 | In one humanoid evaluation, pick-up success was 68.4% from a table, 45.7% from the floor and 76.3% from a shelf. | Google DeepMind |
Industry research we follow
-
NVIDIA16 and 18 March 2026
Describes collecting demonstrations through teleoperation, then expanding them with augmentation, synthetic data and evaluation. Real and simulated data appear as parts of one workflow.
A vendor’s description of its own tools and their intended benefits. NVIDIA blog
-
OpenAIJob posting, checked October 2026
A Robotics Software Engineer posting describes expanding a robotics data collection and evaluation programme, including collection labs and quality control processes.
Evidence of hiring plans. It does not show that OpenAI buys data from outside suppliers. OpenAI careers
-
Google DeepMind30 July 2026
Gemini Robotics 2 adds whole-body control and dexterous manipulation across several robot types. Google notes that multi-finger dexterous manipulation remains challenging, and its results differ by task and setting.
Results from specific evaluations, which do not show general human-level reliability. Google DeepMind
-
Morgan Stanley Investment ManagementJanuary 2026
Identifies large-scale, high-quality, human-centred data as a common bottleneck for embodied AI, and describes teleoperation and motion capture as slow and expensive ways to collect it.
An investment view. Its adoption timing is a forecast and may change. Morgan Stanley (PDF)
-
Goldman SachsFebruary 2024
Forecasts a $38 billion humanoid robot market by 2035 and 1.4 million shipments, linking the revision to AI progress and falling hardware costs.
A forecast of robot sales. It does not measure spending on training data or the market NovaCrest serves. Goldman Sachs
Sources and licence
Anthropic, What work can robots do, and the robot exposure dataset documentation. U.S. Department of Labor, O*NET 29.3.
Anthropic data and O*NET 29.3 content are used under CC BY 4.0. O*NET is a trademark of USDOL/ETA, which has not approved, endorsed or tested these modifications. NovaCrest works on robotics data operations, and this analysis gives context rather than a data-purchase recommendation.
Partnerships
Collection where the gaps are
Tell us which tasks and settings your robots need to learn, and we will plan collection around them.