Socially-Weighted Distributed Graph Optimization

Autonomous service fleets—retail carts, clinical pushchairs, airport trolleys—must route continuously through crowded public space, where classical planners treat pedestrians as physical obstacles rather than people whose personal space carries a cost. We propose Distributed Dynamic Route Optimization (D²RO), powered by Socially-Weighted Distributed Graph Optimization (SW-DGO): a multi-objective cost field minimised subject to directional corridor reservations and solved via incremental D* Lite graph repair over V2V mesh telemetry.

Across 4,650 trials in supermarket, hospital, and airport topologies, proxemic routing eliminates intimate-space exposure (reducing median exposure to 0.00 person-seconds). Controlled experiments establish when distributed coordination is warranted: mesh telemetry and corridor reservations earn their cost under hidden blockages and single-file bottlenecks, advancing rerouting by 11.1 s and raising corridor mission success from 36% to 88%.

github.com/polla-fattah/SW-DGO  —  source, datasets and tests

Code under MIT, datasets under CC BY 4.0.


Contents & Research Assets


Problem Context

Classical planners fail here in two specific ways. A shortest-path planner drives through a browsing shopper's personal space because nothing in its objective says not to, and reactive avoidance only engages once someone is already close — by which point the intrusion has happened. Separately, two carts entering one single-file aisle from opposite ends is not a collision-avoidance problem but a coordination problem, and a purely local controller can only oscillate.

Experimental Results

Planner Success Makespan (s) Exposure (person-s, median)
SW-DGO Proposed 99.0% 47.18 ± 13.40 0.00
Local Social D* Lite 100.0% 39.06 ± 15.12 0.00
Static A* (matched controller) 100.0% 19.20 6.40
Static A* 100.0% 18.00 6.40
Artificial potential fields 100.0% 34.54 10.18

The table above is the broad randomized benchmark, and the proposed system does not win it. Local Social D* Lite — the same proxemic routing with no mesh and no reservation — matches its social compliance exactly and finishes about 8 s sooner. That 8 s is the overhead of a distributed layer whose target topologies a randomized supermarket run rarely produces.

The two experiments below construct those topologies deliberately. Both are N = 50 seed-paired trials with the mechanism the only thing switched.

Controlled mechanism experiment Mechanism ON OFF p
Anticipation lead time (s) V2V mesh 10.70 ± 4.20 −0.10 ± 0.04 3.8 × 10−9
Backtracking distance (m) 1.08 ± 0.68 2.73 ± 0.87 3.2 × 10−8
Corridor mission success Reservation 88.0% 36.0% 1.7 × 10−4
Off-corridor vertices (diversion) 2.16 ± 1.36 0.00 2.3 × 10−8

Read together, the two tables give the actual finding, which is conditional rather than promotional: the distributed layer is not a general improvement but a targeted one. A deployment whose topology routinely produces blockages beyond sensing range, or single-file corridors entered from both ends, should pay for the radio. One that does not should run the local planner — simpler, faster, and equally well behaved around people.

The reservation also turns out not to work the way its name suggests. Head-on encounters are statistically unchanged and total waiting is 0.03 s: agents do not queue for the corridor, they leave it. We therefore describe it as cost-projected diversion rather than mutual exclusion.

A 2×2 factorial with the mesh and reservation disabled in every cell isolates the proxemic cost term: enabling it alone reduces exposure by 6.40 person-seconds (95% CI [−6.48, −6.31], p = 1.5 × 10−9), at a cost of 19.03 s of travel time. Reactive yielding adds no statistically detectable reduction on top of it (p = 0.59).

The distributed layer is conditional. An ordinary human-aware planner with no mesh and no reservation matches that social compliance exactly (p = 1) while finishing about 8 s sooner. The mesh and reservation earn their cost only under the topologies they were designed for — a blockage outside a follower's sensing range, and a single-file corridor entered from both ends.

Limitations & Scoping

  • Simulated pedestrians do not react to the robots; reported exposure is an upper bound on a real crowd.
  • Weights are hand-selected. The sensitivity sweep shows a plateau but claims no optimality, and is one-at-a-time.
  • Sensing is range-limited, without ray-casting against shelf occlusion.
  • Fleet scaling is bounded: success falls to 78% at twelve carts in 36 × 24 m.
  • ORCA and local MAPF are our own implementations and complete 0% of missions; no conclusion depends on them.
  • All results are simulation. No physical fleet was deployed.
  • We report when the distributed layer helps, not how often those conditions arise.

Future Research Directions

  • Joint Weight Calibration: Multi-objective Pareto optimization over travel makespan and social exposure, or automated weight selection replacing one-at-a-time sensitivity sweeps.
  • Mesh Traffic Suppression: Spatial scoping and relevance filtering of V2V broadcasts to scale beyond the 12-cart fleet density ceiling.
  • Heading-Augmented Search: Incorporating turn penalties (αturn) directly into graph search over an orientation-augmented state space rather than relying solely on motion-layer kinodynamics (ωmax = 2.5 rad/s).
  • Hybrid Learning-Guided Search: Coupling Graph Neural Networks or deep RL (e.g., PRIMAL / Learn to Follow) for global goal allocation with the deterministic D* Lite engine.
  • Heterogeneous Multi-Agent Ecosystems: Extending SW-DGO to mixed fleets (shopping carts, clinical pushchairs, floor scrubbers, wheelchairs) via vehicle-specific agility weights in directional corridor reservations.
  • Physical Deployment & Localization: Validating RF attenuation in real metal-gondola retail environments and fusing UWB / IMU / wheel odometry through EKF localization.

Publication Status: A manuscript describing this work is under peer review. Citation details will be added on acceptance.