Most conflict-resolution systems for Advanced Air Mobility (AAM) handle long-range planning and in-the-moment maneuvering as two separate systems built independently. This approach can be risky because these systems can end up telling the same aircraft to do conflicting things.
The framework in our technical paper, “Learning-based Hierarchical Framework for Strategic and Tactical Planning,” takes a different approach. Drawing on NASA’s Stratway work, which decouples trajectory generation and conflict detection from resolution specifically to make verification and certification tractable, we introduce a modular architecture that brings strategic and tactical deconfliction under one system with continuous verification between the layers. Instead of building one monolithic conflict resolution system, each layer is its own module, and those modules can be swapped out to match different use cases.
The Problem with How CDR is Usually Built
Industry surveys of Conflict Detection and Resolution (CDR) research find that tactical, distributed solutions are optimized for fast, local avoidance. That focus leaves a critical gap: many systems treat strategic and tactical deconfliction as independent processes. When both run simultaneously without coordination, their advisories can conflict with each other or create deadlocks.
Other issues in the field include stochastic events (weather, sensor noise, communication drops) that tend to get handled heuristically rather than through a consistent method across planning layers, certification pathways for learned (ML-based) controllers that remain underdeveloped, optimization-based approaches that don’t scale well computationally as traffic density grows, and the absence of a standardized benchmark for evaluating strategic-tactical systems against each other.
As air traffic grows and UAS and eVTOL operations increase, closing this gap becomes more urgent. Using a tree-based strategic planner, which searches through sampling rather than exhaustively evaluating every possible action, offers a scalability and efficiency advantage over baseline approaches that solve conflict detection and resolution in a monolithic, end-to-end way while simulating a large number of aircraft. Breaking the system into modules is about handling that growing traffic density as much as it is about certification.
CDR Architecture
The framework splits into two top-level pieces. A Deconfliction Monitoring Module (DMM) handles detection, using a trajectory forecaster to project future aircraft states and a conflict detector to flag problems, whether that’s a demand-capacity imbalance, a loss of separation between aircraft, or a corridor or vertiport overload.
Detected conflicts get passed to the Conflict Resolution Module (CRM), which is organized into three tiers by time horizon:
- Strategic Planner: operates minutes to hours ahead of time, before aircraft even depart, working across the full state and intent of all traffic to balance demand and capacity and enable long-horizon deconfliction. It’s implemented as a Variable-Step Single-Player MCTS. One search tree plans joint actions for all aircraft at once, and it branches whenever any single aircraft reaches a decision point rather than waiting for all of them, which lets it plan for aircraft that start and finish maneuvers at different times.
- Tactical Planner: operates on a minutes-long horizon, in-flight, working only with the aircraft directly involved in a specific conflict, adjusting velocity or altitude as needed.
- Escape Maneuver: a last-resort, seconds-long layer meant to run on preset rules and lookup tables. The paper defines this layer architecturally but doesn’t implement it separately in the current work. The tactical planner currently covers this range as well.
A routing mechanism ties the layers together, and conflicts get sent to whichever tier matches their time-to-conflict threshold. That routing runs on a continuous verification loop, which is what keeps the strategic and tactical layers from ever issuing contradictory advisories to the same aircraft.
Modularity Going Beyond the Tactical Layer
Any module in this architecture can be swapped out depending on the use case. In this paper, the tactical layer was tested with two different resolvers: a rule-based deterministic one and a multi-agent reinforcement learning (MARL) one, both plugged into the same strategic layer without any other changes. The same framework would support swapping in a different strategic planner as well. This lets us certify parts of the system individually, rather than certifying a single large black box.
The Results
The framework was tested against three stress cases in a simulated UAM network over the New York metropolitan area, with four vertiports placed at the real coordinates of JFK, LaGuardia, Teterboro, and Newark, connected by six corridors. Each case was also run against two baselines: no intervention at all and a reactive tactical-only system with no forecasting or strategic planning. The results support the architecture claim:
- An overloaded vertiport and corridor: with eight aircraft competing for a vertiport and corridor that could each handle three, the strategic planner cut capacity violations from 52 to zero before any aircraft departed, using just two departure delays and one reroute, and every aircraft still reached its original destination. The reactive baseline couldn’t resolve a single violation, because capacity conflicts have to be solved before departure.
- In-flight conflicts between converging aircraft: the tactical planner resolved the conflicts with zero unsafe encounters, keeping aircraft roughly 2,000 meters apart at their closest, versus about 1,000 meters for the reactive baseline and 190 meters (an actual loss of well-clear) with no intervention. Forecasting conflicts ahead of time also meant fewer corrections: four speed advisories instead of the reactive baseline’s six.
- An unexpected corridor closure: the strategic planner rerouted around the blockage, the tactical planner cleaned up the resulting conflicts from that reroute with just two speed adjustments, and once again every aircraft reached its original destination with zero remaining capacity violations, whereas the reactive baseline, unable to reroute, ended up with eleven.
The two swappable tactical resolvers reached the same safety outcome but with different tradeoffs. The MARL-based resolver was measurably more efficient, resolving the same conflicts with 16.7% fewer advisories, while the rule-based resolver is deterministic and easier to certify.
A modular architecture like this lets operators adapt the framework to their own operational needs, choosing the right resolver for the job as AAM traffic scales up. Read the paper to learn more.