Five researchers, four at the University of California, Irvine and one at the Indian Institute of Technology Bombay, posted CORNAV to arXiv on October 2, 2026. It is a navigation system for site robots that works from two documents a general contractor already keeps, the 2D CAD floor plan and the weekly look-ahead schedule. Across tests at an active construction site and an office, the full system completed 72.2% of navigation tasks, against 13.0% for the same system with the floor plan taken out. Genova Construction supported the work in part and provided the site and one of the two robots.
What it does
A person gives the robot a plain-language request with a date and time, such as “find the chair in office 1 on floor 0,” and CORNAV returns either a route or a refusal with a reason. Setup runs once per site, from three inputs:
- The floor plan, exported from CAD as a DXF file. The software reads each room as a closed polyline and aligns the plan with a map the robot built while walking the site.
- The look-ahead schedule, kept as a spreadsheet. Each activity carries a room, an active time window, hazard notes, and a severity of hard or soft. A hard zone is closed to the robot, for example during hazardous work or over wet concrete. A soft zone stays open but costs more to cross.
- Camera and depth data from a first walk of the site, turned into a scene graph, which is a list of the objects the robot saw and where they sit.
When a request comes in, the schedule decides which zones are active at that hour. A large language model, the kind of AI behind ChatGPT (GPT-4o in this paper), reads the hazard notes on each active soft zone and scores them from 0 to 1 against four rules the authors drew from OSHA’s construction standards in 29 CFR 1926. The rules cover temporary structures such as scaffolding, unprotected edges, active work zones, and unfinished surfaces. A score of 0.7 or higher turns the soft zone into a hard one, and a standard shortest-route search plans around everything closed.
On a contractor’s team, the superintendent who writes the look-ahead and the VDC or technology lead who runs a site robot would both work with it. Because the robot takes its closed zones from the look-ahead, an activity with no room assigned gives it nothing to avoid.
Evidence so far
The evidence is the authors’ own testing, reported in the paper. Every quantitative result was computed offline from data a Unitree Go2, a four-legged robot, recorded on three floors of the construction site and one office floor. The authors also ran 12 live missions on the Go2 and a Unitree G1 humanoid and report that the robots followed the schedule and did not move when no allowed route existed.
The authors reran the same 54 requests with one component removed at a time, a test called an ablation, and compared against HOV-SG, the published scene-graph method CORNAV builds on.
| Method | Routes clear of closed zones (%) | Task success (%) | Requests with a feasible route (%) |
|---|---|---|---|
| CORNAV full system | 100.0 | 72.2 | 46.3 |
| Without the floor plan | 98.1 | 13.0 | 40.7 |
| Without the schedule | 53.7 | 72.2 | 40.7 |
| Without the safety check | 83.3 | 72.2 | 46.3 |
| HOV-SG baseline | 42.6 | 7.4 | 31.5 |
Task success depends on the floor plan. The scene graph’s own room labels matched the drawing in 1 of 30 room segments, and in none of the segments on the three construction floors. Without the CAD rooms, the robot found none of the permanent fixtures shown on the drawing.
Removing the schedule moved the compliance column. The share of routes that stayed out of every closed zone fell from 100% to 53.7%, while the planner still returned routes for 40.7% of requests, the same rate as the version without the floor plan. So a count of returned routes says little about whether the robot respected the look-ahead.
In a separate sweep, the authors issued requests at 89 different times across four weekly look-ahead schedules. Of those, 74 had an allowed route, and none of the 74 crossed an active hard zone. The other 15 were refused because the start or the goal sat inside a closed zone (nine starts and six goals).
| Area | Trials | Feasible | Hard-zone violations (%) | Soft zones avoided (%) | Soft zones crossed (%) |
|---|---|---|---|---|---|
| Construction floor 0 | 22 | 20 | 0.0 | 0.0 | 100.0 |
| Construction floor 1 | 23 | 19 | 0.0 | 78.9 | 21.1 |
| Construction floor 2 | 23 | 18 | 0.0 | 0.0 | 100.0 |
| Office | 21 | 17 | 0.0 | 0.0 | 100.0 |
The authors attribute the 100% soft-zone crossing rates to floor geometry, since in those areas the soft zone sat on the only path between start and goal.
The safety check was tested on eight cases the authors wrote, one hazardous and one safe case per rule, with each hazardous zone deliberately labeled soft. It escalated all four hazardous cases and left the four safe ones alone, and scoring each case eight times, 64 calls in all, never changed a decision. Those notes were written to match the rules. Results on hazard notes taken from working look-ahead schedules have not yet been reported.
How to measure it on your projects
Anyway, the metric to borrow is the paper’s own. Zone compliance rate is the share of robot missions whose logged path stayed outside every area the look-ahead had closed at that hour. You check it by overlaying each logged path on that day’s floor plan with the hard zones marked.
Capture the baseline from the schedule before any robot runs. Take the next two weeks of the look-ahead and count the share of activities that name a specific room or area and carry a hazard level. If that share is low, a schedule-aware robot has few zones to avoid, and a 100% compliance rate would mean little.
A first step on one job is to take this week’s look-ahead for a single floor, add a room and a hard or soft tag to each activity, and have the superintendent mark the hard zones on the DXF floor plan. Those are the inputs the paper used.
Availability and cost
CORNAV is a research system with no commercial release. The project page lists it as under review at ICRA, a robotics conference, and shows the abstract, the pipeline diagram and a sample weekly look-ahead schedule. Code and the recorded site data are not posted there. The safety check in the paper runs on GPT-4o.