CORNav combines construction drawings, a project schedule, site-observed objects, and explicit safety rules so a robot can choose routes that change as work zones open and close. In recorded-data tests across 89 schedule-time trials, it produced no paths through hard-restricted zones, and it rejected all 15 requests whose start or destination lay inside one—addressing a key weakness of navigation systems that treat worksites as static maps.
What Did the Researchers Build?
CORNAV is a navigation system designed for construction sites, where room layouts, work activities, and hazards change over time. Instead of relying on a building information model (BIM), it brings together the CAD drawings and project schedules that construction teams already maintain, along with a robot’s observations of the site.
The system uses CAD drawings as its spatial reference: they define rooms, their boundaries, and the locations of permanent fixtures. A spreadsheet project schedule is converted into structured information about activities, assigned rooms, active time windows, and hazard severity. Separately, RGB-D observations—color images paired with depth measurements—from an initial site survey are used to build a 3D scene graph: a representation of objects and their spatial relationships.
At run time, CORNAV combines these sources with a task request. It can find movable objects seen during the survey and permanent fixtures recorded in the drawings, then associate them with rooms. It also turns scheduled work and safety information into areas the robot must avoid or should avoid. This combination is intended to help navigation account for both what is physically present and what is planned.
What Were the Key Results?
The schedule-aware evaluation tested 89 distinct schedule timestamps drawn from four weekly look-ahead project schedules. Once active hard constraints were applied, 74 requests had feasible routes. The other 15 were correctly reported as infeasible because either the start or the goal was inside a restricted area—not because the route between accessible endpoints had been cut off.
| Schedule-sweep result | Reported outcome |
|---|---|
| Trials across schedule timestamps | 89 |
| Trials with a feasible route | 74 |
| Trials correctly reported infeasible | 15 |
| Hard-zone violations on feasible routes | 0 |
| Infeasible trials with goal inside a hard zone | 6 |
| Infeasible trials with start inside a hard zone | 9 |
The team checked every waypoint on every planned route against the active hard-zone boundaries. None of the waypoints on feasible routes fell inside a restricted area. This is a direct test of constraint compliance, rather than a general claim that the routes were safe in every respect.
Path efficiency—the planned path length divided by the path length without schedule constraints—had a mean near 1.0 across environments. Some floor-one routes lengthened to avoid active cautionary areas, while variation on floor zero was attributed to detours around narrow passages. The available results do not provide a precise mean or a comparison against a competing system.
How Does CORNAV Work?
CORNAV separates preparation from navigation. Before tasks begin, the system aligns CAD-derived room boundaries with the robot’s map. It organizes the schedule into activities with room assignments, active times, hazard notes, and severity. A site survey supplies the 3D scene graph, which stores observed objects, their relationships, and language-related features.
When a task arrives, CORNAV resolves its target using the data source best suited to that object. For movable objects observed during the survey, it compares the task’s text embedding with stored visual embeddings using cosine similarity—a measure of how closely two representations match. For permanent fixtures listed in CAD, such as toilets, it looks up the fixture in a drawing-derived registry instead. The associated room is determined by checking whether the object’s position falls inside an aligned room boundary.
Next, the system checks the schedule for the requested execution time. An active activity becomes a geometric region in the robot’s coordinate frame. Hard constraints—such as active hazardous work or wet concrete—are added to the occupancy grid as occupied cells. Since the route search treats them like obstacles, a valid route cannot cross them. If a task starts or ends inside one, the planner reports that it cannot produce a feasible route.
Soft constraints work differently: the region remains traversable, but crossing it carries an extra cost, encouraging the planner to take another route when practical. A safety-validation module checks schedule hazard information against rules covering temporary structures, fall hazards, active work, and unstable or unfinished surfaces. It can escalate a cautionary region to a hard exclusion zone.
Why Does This Matter for Robotics?
A construction site is not a finished building with a stable map. A corridor can be clear at one time and restricted later; a room may contain an active work crew or an unfinished surface. A route that was valid during one shift can therefore become unsafe when the schedule changes. CORNAV’s approach makes planned work—not just objects detected by sensors—part of the robot’s navigation decisions.
The use of CAD drawings and spreadsheets is also practically relevant. The system does not require a BIM or IFC model, and it can connect work activities to drawn room boundaries. That could make schedule-aware navigation more accessible on projects where those structured building models are unavailable or incomplete. Its open-vocabulary object grounding also lets task requests refer to movable items observed on site, not only objects listed in a building model.
The paper reports quantitative tests on recorded sensor data and qualitative demonstrations of the full pipeline on a Unitree Go2 quadruped and a G1 humanoid. That makes the work relevant to teams exploring mobile robots for site inspection, material movement, or other construction tasks, although the results do not establish readiness for unsupervised operation. For teams evaluating humanoid platforms, browse humanoid robots on Robot Overflow.
What Are the Limitations and Open Questions?
The quantitative navigation results were computed offline using RGB, depth, and pose data recorded during site surveys. The paper also reports live demonstrations on two robot platforms, but describes those demonstrations qualitatively; the schedule-sweep numbers should not be read as live, on-site performance figures.
CORNAV also depends on aligned drawings, usable project schedules, and an initial site survey. If room assignments or hazard notes are missing or wrong, the system’s decisions may be affected. The safety rules cover important construction risks, but the reported tests do not establish that the list covers every hazard or that schedule validation catches every annotation error. Further evaluation would need to measure performance across more sites and changing conditions, and quantify how often the safety-validation module correctly handles mislabeled hazards.
Frequently Asked Questions
What problem does CORNAV address? It plans robot routes on construction sites where scheduled work can make areas temporarily unsafe. It distinguishes areas that must be avoided from cautionary areas that should be avoided when possible.
Does CORNAV require a BIM model? No. It uses CAD drawings and project schedules, alongside a 3D scene graph built from site observations.
What happens if a task starts or ends in a restricted zone? The planner reports the task as infeasible rather than returning a route through a hard-restricted area. In the schedule-sweep evaluation, all 15 such cases were identified.
Was the system tested on physical robots? The full pipeline was demonstrated qualitatively on a Unitree Go2 quadruped and a G1 humanoid. The reported schedule-sweep measurements were calculated offline from recorded data.
CORNAV shows how construction drawings, work schedules, and robot perception can be combined to make navigation respond to changing site conditions. Its results demonstrate strict avoidance of the hard zones in the reported schedule-sweep tests, while leaving broader live performance and hazard coverage as open questions.