A preprint proposes a vocabulary for describing what construction workers do, with the aim of giving robot builders a more precise inventory of activities and required capabilities. Called TARCAT, the system defines 41 reusable primitives, or basic building blocks for describing work, across intellectual, social and physical categories. It was derived from 91 O*NET task statements across seven high-employment construction occupations and 30 instructional videos. The authors position it as a common vocabulary for organizing demonstrations and specifying robot requirements.
Starting with the work itself
To define the occupational frame, the authors used 2024 employment data and selected the seven construction occupations with the largest number of employed workers. They then drew 169 task statements from O*NET, an occupational database. Of those statements, 137 involved physical movement and 32 were non-movement. The source material was a set of occupational records rather than a participant cohort.
Video collection was selective. ChatGPT with GPT-5.6 was used to search YouTube, and 30 instructional videos were manually retained. They covered 59 movement tasks. The 32 non-movement tasks were annotated from the occupational statements, giving reported coverage of 91 O*NET tasks overall. Manual retention means the set is a curated foundation for the taxonomy; the supplied analysis does not establish that it represents construction activities broadly.
Turning tasks into building blocks
The important shift is from a job description to a reusable unit. TARCAT represents each activity as a context-specific instantiation of a primitive, with parameters that can identify the object, tool, target or other details. In practical terms, the primitive is the reusable pattern, while the parameters describe the situation in which it is used. The paper defines a skill as a capability acquired through learning or programming to execute an activity or sequence toward a goal, and says lower-level skills can be recursively composed into higher-level skills.
Time-stamped annotation of the videos yielded 643 primitive labels. Arm translation and positioning was the most frequent subcategory, with 278 occurrences, followed by problem solving with 114. Those figures describe the labels in this annotated material. They do not estimate how well a robot performs either capability.
A small step onto hardware
To show how the vocabulary might connect to a machine, the authors selected four physical activities, mapped them to primitives and scripted them on a six-axis DOBOT CR3 arm with a 15-actuator, tendon-driven CRAFT hand. The demonstrations covered arm translation and positioning, arm and wrist rotation, and surface-contact manipulation. The authors describe this as initial evidence of executable physical instantiation.
That hardware step remains narrow. The study reports four selected scripted activities, while its taxonomy contains 41 primitives and covers 91 O*NET tasks, and it provides no reported comparator or success metric for the demonstrations. The occupational scope also excludes heavy machinery operators because that occupation was outside the study scope. These limits leave open whether the taxonomy transfers to longer, more varied construction work.
The authors identify long-horizon skills and data-centric primitive classification as future work. Those directions matter because the current study organizes tasks and demonstrates selected primitives, but does not report a long-horizon skill evaluation or a data-centric classification benchmark. The paper states that its annotations are available through the TARCAT GitHub repository and that the physical demonstration video is available online. The supplied document is an arXiv preprint, version 3, dated 24 September 2026.
Paper data and sources
Original title: A Taxonomy of Construction Task Activities for Robot Workers
Authors: Sadman Sakib, Zhangyi None Peng, Yujie Pang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text