CORNAV: Construction-Aware Reasoning for Robot Navigation on Active Worksites
This preprint presents CORNAV, a navigation system that lets robots operate on active construction sites by combining 2D CAD blueprints, project schedules, open-vocabulary 3D scene graphs, and LLM safety checks. In ablation tests, blueprint grounding raised task success from 13.0% to 72.2% and schedule awareness eliminated hard-zone violations, while safety validation cut hazardous-zone entry from 100% to 0% for mislabeled hazards. The approach matters because it avoids the need for BIM models and lets robots follow site safety constraints that change over time.
Can a robot navigate active construction sites reliably by grounding natural-language task requests in 2D CAD drawings and project schedules, without requiring a Building Information Model, while enforcing schedule-dependent safety constraints?
Construction sites change continuously, with temporary work zones, partial structures, and safety exclusions that can shift by day and time. Existing language-grounded navigation systems rely on semantic scene understanding alone and lack architectural plans, project schedules, and safety rules, so they localize permanent building features unreliably and cannot safely adapt to evolving jobsite hazards. BIM-based alternatives require costly IFC models that many contractors do not maintain and treat hazards as static obstacles.
Prior work includes open-vocabulary 3D scene graphs such as Hydra, OVSG, ConceptGraphs, and HOV-SG, which enable language-based object retrieval but often mislabel rooms and do not incorporate schedule or safety constraints. BIM/IFC-based systems by Follini et al., Gopee et al., Karimi et al., and others can produce schedule-aware maps or routing costs, but they presuppose a BIM or IFC model, treat declared hazards as fixed exclusion zones, and resolve queries through structured class lookups rather than open-vocabulary grounding. Amani et al. adjust navigation from prompt sentiment but do not validate hazards against formal safety rules.
CORNAV aligns DXF CAD floor plans to robot occupancy maps using a two-stage rigid alignment: coarse rotational search over rasterized walls followed by ICP refinement. It builds a blueprint scene graph for deterministic room resolution and stores permanent architectural fixtures directly from CAD. Offline, RGB-D survey data are processed with HOV-SG to build an open-vocabulary 3D scene graph for CLIP-based retrieval of observed objects. Project spreadsheets are converted to time-varying constraints with hard and soft hazard levels. During each query, an LLM (GPT-4o) scores active soft zones against OSHA-derived safety rules and promotes zones scoring at least 0.7 to hard constraints. An A* planner uses an obstacle-clearance cost field, with hard zones rasterized as occupied cells and soft zones given a traversal-cost multiplier of 5, returning waypoints for ROS 2 execution.
Blueprint-to-occupancy-map alignment RMSE was 0.135 m in the office and 0.280–0.386 m across three construction floors. HOV-SG room-label accuracy against blueprint ground truth was 3.3% overall (0% on construction floors, 16.7% in the office). In the main ablation over 54 navigation requests, CORNAV achieved 100% constraint compliance, 72.2% task success, 46.3% feasible planning rate, 29.6% schedule-aware compliance rate, and 0.988 path efficiency ratio. Removing blueprint grounding dropped task success to 13.0%, with 0% on permanent objects and 17.1% on CLIP objects; removing schedules dropped compliance to 53.7%; removing safety validation dropped compliance to 83.3%. The HOV-SG baseline achieved 7.4% task success, 42.6% compliance, 31.5% feasible planning, and 1.398 path efficiency. In 89 schedule-aware trials across four sites, 74 were feasible and none violated hard zones; 15 were correctly reported infeasible because start or goal fell inside an active hard zone. Safety validation experiments over 8 cases (4 hazardous, 4 non-hazardous) escalated all hazardous soft zones to hard and left all non-hazardous zones soft; hazardous-zone entry rate fell from 100% to 0%, and scoring was stable with only one case showing standard deviation 0.024 across 8 repeats. Twelve live qualitative trials on a Unitree Go2 quadruped and G1 humanoid reportedly reached destinations or correctly refused infeasible paths.
The authors note future work should address multi-robot coordination and coupling navigation with mobile manipulation. Quantitative evaluation was offline on recorded Go2 data, with live deployment only qualitative and limited to 12 trials at two sites. Room-label accuracy and alignment RMSE depend on map quality and CAD completeness. Safety validation relies on GPT-4o and a fixed 0.7 threshold, and schedule quality is assumed. The paper does not report code release or peer review, and all experiments use a single robot survey with 54 planning requests and 8 safety cases, so broader generalization is not established.
Construction robotics integrators, mobile robot OEMs such as Unitree, and AEC software providers could use this to add schedule-aware, CAD-grounded navigation to inspection, material transport, or site monitoring robots without requiring BIM adoption. The approach is closest to deployable now on sites that already maintain 2D CAD drawings and scheduled work zones; within 1–3 years, it could become a differentiated feature for construction robot fleets if safety rules are adapted to regional regulations and alignment robustness improves. Longer term (3+ years), coupling with manipulation and multi-robot coordination would be required for broad autonomous construction workflows.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.