Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Autonomous Decision Making for Air Taxi Networks

Alex Vesel

TL;DR

The paper tackles scalable, safe autonomous control for dense urban air taxi networks by formulating the Air Taxi Network Problem (ATNP) as a multiagent MMDP and proposing a three-phase solution: agent-passenger assignment, flight level selection, and trajectory planning. It decomposes the problem to manage computational complexity, employing Murty's algorithm for candidate matches, a density-based flight-level risk model, and Monte Carlo Tree Search to refine trajectories, with a centralized ADS-B-enabled simulator evaluating performance on Bay Area and NYC vertiport layouts. Key contributions include the ATNP-MMDP formulation, a practical three-phase solver, and a simulation framework demonstrating improvements in safety and passenger wait times versus greedy and first-dispatch baselines. The work advances autonomous, high-density UAM by providing a tractable planning pipeline and empirical evidence of safety-throughput gains, informing future ATM design for UML5-scale systems.

Abstract

Future urban air mobility systems are expected to be operated by rideshare companies as fleets, which will require fully autonomous air traffic control systems and an order of magnitude increase in airspace capacity. Such a system must not only be safe, but also highly responsive to customer demand. This paper proposes the air traffic network problem (ATNP), which models the optimization problem of future cooperative air taxi networks. We propose a three-phase decision making model that efficiently assigns vehicles to passengers, determines flight levels to reduce collision risk, and resolves aircraft conflicts by selectively applying Monte Carlo tree search. We develop a simulator for the ATNP and show that our approach has increased safety and reduced passenger waiting time compared to greedy and first-dispatch protocols over potential vertiport layouts across the Bay Area and New York City.

Autonomous Decision Making for Air Taxi Networks

TL;DR

The paper tackles scalable, safe autonomous control for dense urban air taxi networks by formulating the Air Taxi Network Problem (ATNP) as a multiagent MMDP and proposing a three-phase solution: agent-passenger assignment, flight level selection, and trajectory planning. It decomposes the problem to manage computational complexity, employing Murty's algorithm for candidate matches, a density-based flight-level risk model, and Monte Carlo Tree Search to refine trajectories, with a centralized ADS-B-enabled simulator evaluating performance on Bay Area and NYC vertiport layouts. Key contributions include the ATNP-MMDP formulation, a practical three-phase solver, and a simulation framework demonstrating improvements in safety and passenger wait times versus greedy and first-dispatch baselines. The work advances autonomous, high-density UAM by providing a tractable planning pipeline and empirical evidence of safety-throughput gains, informing future ATM design for UML5-scale systems.

Abstract

Future urban air mobility systems are expected to be operated by rideshare companies as fleets, which will require fully autonomous air traffic control systems and an order of magnitude increase in airspace capacity. Such a system must not only be safe, but also highly responsive to customer demand. This paper proposes the air traffic network problem (ATNP), which models the optimization problem of future cooperative air taxi networks. We propose a three-phase decision making model that efficiently assigns vehicles to passengers, determines flight levels to reduce collision risk, and resolves aircraft conflicts by selectively applying Monte Carlo tree search. We develop a simulator for the ATNP and show that our approach has increased safety and reduced passenger waiting time compared to greedy and first-dispatch protocols over potential vertiport layouts across the Bay Area and New York City.
Paper Structure (27 sections, 3 equations, 3 figures, 3 tables, 3 algorithms)

This paper contains 27 sections, 3 equations, 3 figures, 3 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Vehicle Routing Problem
  4. Aircraft Collision Avoidance
  5. Rideshare Matching
  6. Problem Statement
  7. Agents
  8. Vertiports
  9. Passengers
  10. ATNP as a Multiagent Markov Decision Process
  11. Method
  12. Agent-Passenger Assignment
  13. Candidate Matching Generation
  14. Final Matching Selection
  15. Flight Level Selection
  16. ...and 12 more sections

Figures (3)

  • Figure 1: A high level system architecture diagram. At each timestep, the fleet of eVTOLs broadcast their position and heading information to the controller via ADS-B. The controller computes a joint action and broadcasts back to each eVTOL. Each eVTOL takes the joint action over a timestep, resulting in a new map. This process repeats until all passengers are delivered.
  • Figure 2: An example of aircraft density grids used in flight level selection for 1, 10 and 20 steps into the future. Note that as the number of steps increase, the densities becomes more diffuse to account for uncertainty in the future locations of each aircraft.
  • Figure 3: Generated maps for the Bay Area and New York City. Red dots indicate vertiports and blue dots show the tracks of an agent's flight path over 60 seconds. Green dots indicate a landed agent. Note the area of the Bay Area map is larger than NYC.