OmniPath: A Multi-Modal Agentic Framework for Auditing Wheelchair Accessibility¶

🕒 Published (v1): 2026-06-23 04:23 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

OmniPath is an agentic auditing framework that fuses OSM vector topology with USGS 3DEP LiDAR point clouds to proactively detect wheelchair accessibility barriers at 0.5 m resolution, before a user ever traverses a route. It replaces reactive crowdsourced reporting with an autonomous agent that quantifies running slope, cross slope, and surface discontinuities against ADA thresholds, assigning a weighted severity score to every path segment. Validated against 200 field surveys at the National Mall, it achieves F1 = 0.60/0.58 for Severe/Critical hazard categories.

Problem¶

Standard digital maps (OSM, DC-GIS) encode pedestrian network connectivity but not surface quality, creating "Accessibility Traps" where wheelchair-valid-looking routes are physically impassable (e.g., staircases encoded as sidewalks). Crowdsourced approaches (Wheelmap) suffer from data sparsity — unsurveyed barriers remain invisible until a user discovers them physically. No prior system combined high-resolution LiDAR with an agentic AI pipeline specifically for ADA-compliance auditing at the micro-scale relevant to wheelchair mobility.

Method¶

Data fusion: OSM pedestrian graph (1,485 segments, National Mall) is geospatially aligned to 2021 DC 3DEP LiDAR (QL1/QL2, ≥8 pulses/m², ±5 cm vertical) via coordinate reference system normalization.

Micro-scale segmentation (Agent A): Every OSM edge is discretized into non-overlapping 0.5 m sliding windows (approximating wheelchair footprint). For each window, a spatial query retrieves LiDAR ground-classified points (class 2) within a 0.5 m perpendicular buffer, projected onto the along-path axis and binned at 10 cm intervals using median elevation.

Feature extraction: - Running slope: linear regression over the elevation profile, $\text{slope} = \frac{\Delta z}{\Delta x} \times 100\%$ - Cross slope: elevation gradient perpendicular to centerline - Discontinuity: abrupt vertical change $> 2$ cm between consecutive 10 cm bins

A second 2.0 m / 1.0 m stride pass detects macro-features (stairs: 8–25 cm risers, 15–45 cm treads; landings: slope < 1%; overhead structures: bimodal LiDAR elevation distribution).

Weighted Severity Score: Proportional exceedance $E = \max\!\left(0, \frac{\text{measured} - T}{T}\right)$ per metric; then: $$S = (E_{\text{running}} \times 4.0) + (E_{\text{cross}} \times 2.5) + (E_{\text{disc}} \times 3.5)$$ Segments classified: Mild ($S < 0.5$), Moderate ($0.5 \leq S < 1.5$), Severe ($1.5 \leq S < 3.0$), Critical ($S \geq 3.0$). Impedance assigned per class feeds the routing engine.

Key Contributions¶

Proactive agentic auditing paradigm: autonomous virtual traversal replaces human-in-the-loop reporting
Multi-modal fusion pipeline: OSM topology + USGS 3DEP LiDAR at 0.5 m granularity
ADA-compliant anomaly detection with a biomechanically-weighted severity scoring formula
Macro-feature detection pass for stairs, landings, and occluded/overhead sidewalk structures
Empirical validation against 200 field surveys; first system to report LiDAR-based wheelchair accessibility metrics at this granularity for the National Mall

Results¶

303,361 micro-scale audit windows evaluated across 1,485 OSM segments
97/1,485 segments (6.53%) had sufficient LiDAR density for analysis; 93.47% excluded due to tree canopy occlusion or flight-path gaps
100% of analyzable segments contained ≥1 ADA violation; 97.9% (95/97) classified Severe or Critical
331 total discrete violations detected across 97 segments; mean 3.41 violations/segment
234 critical violations identified among those 97 segments
F1-score: 0.60 (Severe), 0.58 (Critical) vs. 200 physical ground-truth field surveys

Limitations¶

Severe LiDAR coverage gap: 93.47% of segments excluded due to tree canopy occlusion and flight-path gaps — system cannot audit the majority of the network
Ground-truth validation limited to 200 surveys at one site (National Mall, Washington D.C.); generalizability undemonstrated
Severity weight coefficients (4.0, 2.5, 3.5) are hand-tuned from ADA guidance and field observation, not learned; may require per-user or per-wheelchair-type calibration
LiDAR data is biennial at best; dynamic barriers (construction, debris, parked vehicles) are invisible to the system
No real-time or on-device inference; framework is batch/offline

Relevance to Harnesses / Meta-Harnesses¶

OmniPath exemplifies a domain-specific agentic harness: it orchestrates a pipeline of specialized sub-agents (geometric perception Agent A, Accessibility Compliance Engine) that each consume the output of upstream steps, echoing the harness-of-agents pattern common in meta-harness research. The two-pass architecture (0.5 m micro-scale → 2.0 m macro-scale) mirrors multi-resolution chaining seen in LLM meta-harnesses, where coarse and fine agents operate on the same substrate at different granularities. The severity scoring formula functions as a structured inter-agent contract — the upstream geometry agent produces normalized exceedance values that the downstream routing agent consumes without needing to re-parse raw LiDAR — a clean separation of concerns that meta-harness designers should note. For researchers building harnesses over sensor/geospatial pipelines, this is a concrete example of how agent boundaries, data contracts, and impedance-weighted routing logic can be formalized in a non-LLM domain.