⚡ electrical engineering

SuperFALCON+: Complexity-Aware Global-to-Local Coordination for Resilient MAV Exploration

SuperFALCON+ is a complexity-aware global-to-local coordination framework that integrates FALCON's guidance with SUPER's local execution to enhance resilient MAV exploration in confined environments by dynamically adjusting targets based on spatial constraints and occupancy, thereby significantly improving mapped volume and reducing planning failures compared to baselines.

Original authors: Jiaxing Li, Hanyu Hong, Haiwen Yuan, Yan Meng, Shifan Bi, Huanyu Wu, Haoyu Cai, Liang Ye

Published 2026-10-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiaxing Li, Hanyu Hong, Haiwen Yuan, Yan Meng, Shifan Bi, Huanyu Wu, Haoyu Cai, Liang Ye

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a small drone sent into a dark, cluttered building to map every corner. Its job is to find new, unexplored space and fly there safely. To do this, the drone relies on two distinct ways of thinking. The first is a long-range planner that looks at the big picture, deciding which distant rooms or hidden corridors are worth visiting next. The second is a local pilot that focuses on the immediate moment, constantly checking for walls, furniture, and tight turns to ensure the drone does not crash. In theory, these two minds should work together seamlessly. In practice, however, they often speak different languages. The long-range planner might point toward a useful destination, but the local pilot might find that the path is blocked by an obstacle too close to navigate, or that the destination lies right against a map boundary where the drone's safety calculations break down. When this happens, the drone can get stuck, crash into a wall, or simply give up, leaving parts of the building unmapped.

Researchers at the Wuhan Institute of Technology have developed a new system called SuperFALCON+ to fix this disconnect. Their work addresses a specific problem in autonomous exploration: how to keep a drone moving when the destination chosen by the global planner is difficult for the local pilot to reach. Instead of trying to merge the two planners into one giant brain, which would be incredibly complex and slow, they built a smart mediator between them. This mediator acts as a translator and a safety buffer. It understands the limitations of the local pilot, such as how much space is needed to turn safely or how close the drone can get to a wall before its sensors become unreliable. When the global planner sends a target, the mediator checks if that target is realistic. If the target is too far away or too close to a dangerous edge, the mediator gently nudges the request to a safer, more reachable spot before passing it along. It also keeps a close watch on the drone's progress. If the drone keeps trying to reach the same difficult spot and failing, the mediator recognizes this as a sign of trouble and ensures the system keeps trying new approaches rather than freezing up.

To test this idea, the team ran simulations in three different challenging environments: a complex office with narrow hallways, a maze with sharp, non-standard turns, and a multi-story building with staircases. They compared their new system against two other popular methods. The results showed that the mediated system was significantly more resilient. In the complex office, the new system mapped out 1,804.30 cubic meters of space, while the other methods mapped less than 1,700 cubic meters. More importantly, the new system never stopped prematurely in any of the ten test runs for that environment, whereas the other methods gave up in two to four out of ten attempts. In the maze, the new system mapped 3,680.90 cubic meters and completed all ten trials, while the other methods failed to finish several times. The system achieved this by slowing down slightly, averaging speeds between 1.05 and 1.27 meters per second, compared to the faster but more erratic 1.68 to 1.85 meters per second of the standard method. This slower pace allowed the drone to navigate tight spots without crashing or getting stuck.

The researchers broke down their system to see which part was doing the heavy lifting. They found that removing any single piece of their solution caused the system to fail to complete the maze trials, though some individual runs still mapped significant volumes before stopping. Without the extra safety buffer around the map edges, the drone crashed into boundaries. Without the watchful supervisor that tracks repeated failures, the drone got stuck in loops. Without the target adjuster that moves difficult goals to safer spots, the drone could not find a path. Only when all three parts worked together did the drone successfully complete every single trial, mapping an average of 3,655.4 cubic meters. The team also took their system to the real world, flying a physical drone equipped with a laser scanner and a powerful onboard computer through a campus area filled with trees and branches. Even when the wind caused the drone to shake and its path to wobble near a large tree, the system kept the drone flying and completed the mission.

The core discovery is that resilience in autonomous exploration does not come from making the drone faster or smarter in a single sense, but from managing the handoff between high-level decisions and low-level actions. By acknowledging that a good destination on a map might be a bad destination for a drone in a specific moment, and by adjusting the plan accordingly, the system keeps the mission alive. The authors note that while the system is highly effective in these tests, it still relies on pre-set rules for how much space to reserve and how far to adjust a target. They suggest that future versions could learn these rules on the fly based on how the drone is actually performing. For now, the work proves that a carefully designed interface between a drone's two brains can turn a mission that would otherwise end in failure into one that successfully explores the unknown.

Technical Summary: SuperFALCON+

Problem Statement

In confined environments, autonomous Micro Aerial Vehicle (MAV) exploration often fails due to a mismatch between global exploration intent and local execution constraints. While global planners can identify informative regions (frontiers), local planners frequently struggle to generate executable trajectories for targets located behind narrow turns, near walls, or at map boundaries. The authors observed that in simulations using the FALCON global planner, this disconnect led to specific failure modes: obstacle-penetrating trajectories, repeated optimization failures, and premature mission termination. These failures leave significant portions of the environment unmapped, not because the global planner lacks information, but because the local execution layer cannot satisfy the geometric and kinodynamic constraints imposed by the immediate environment.

Methodology: SuperFALCON+

The paper proposes SuperFALCON+, a complexity-aware global-to-local coordination framework. Rather than unifying the planners into a single monolithic system, SuperFALCON+ maintains a modular architecture that couples FALCON (global, coverage-oriented guidance) with SUPER (local, safety-oriented trajectory execution). The framework introduces a "Mediation Interface" that manages the transfer of targets between these layers through three specific mechanisms:

Boundary-Constraint Coordination (Pre-flight Configuration):
- Mechanism: Before flight, the system configures an additional spatial envelope ( $B^+$ ) around the nominal task region ( $B$ ). This is calculated as a Minkowski sum ( $B \oplus S(\Delta B)$ ) to provide a "spatial reserve."
- Purpose: This reserve accounts for the local planner's (SUPER) requirements for obstacle inflation and boundary queries, ensuring that local map operations have sufficient coordinate support without violating the physical task limits.
Local-Constraint-Aware Target Adjustment (Online Processing):
- Mechanism: This stage combines a Target-Broker module with SUPER's native free-space projection.
  - Target-Broker: Evaluates the span ( $d$ ) between the current vehicle position and the global candidate target ( $p_g$ ). If $d$ exceeds a threshold ( $d_0$ ), the target is adjusted closer to the vehicle ( $e_p$ ) while retaining the original direction. It also enforces a minimum height bound ( $z_{min}$ ).
  - SUPER Projection: The adjusted request is passed to SUPER, which performs a bounded free-space search ( $P_{free}$ ) within a local occupancy map to find a non-occupied, locally admissible goal ( $p_l$ ).
- Purpose: To reconcile global exploration requests with current local occupancy and kinodynamic constraints before trajectory generation begins.
Repeated-Goal-Aware Handoff Supervision (Execution Monitoring):
- Mechanism: A "watchdog" relay forwards target requests from FALCON to SUPER while monitoring the target stream. It records repeated endpoints (candidates that are repeatedly sent without successful execution).
- Purpose: To maintain the continuity of the target handoff and provide cues of limited progress without interrupting the native replanning loops of either FALCON or SUPER. It does not initiate an escape maneuver but ensures the interface remains active during difficult planning events.

Key Contributions

The authors identify the coordination interface, rather than the individual planning layers, as a critical source of execution interruption in confined environments. Their contributions are:

Modular Architecture: A framework that integrates FALCON's global reasoning with SUPER's local execution, explicitly addressing the mismatch in spatial horizons and operating timescales.
Complexity-Aware Coordination: A three-stage response system (configuration, target adjustment, and handoff supervision) that translates operational complexity into specific coordination actions.
Empirical Validation: Integration and evaluation of the system in three confined simulation scenarios and on a physical quadrotor, demonstrating improved resilience compared to baselines.

Experimental Results

The system was evaluated in three confined benchmark environments: Complex Office, Octa Maze, and Duplex Office. Performance was compared against EDEN and FALCON baselines.

Mapped Volume: SuperFALCON+ achieved the highest mean mapped volume in all scenarios (e.g., 1804.30 m³ in Complex Office vs. 1692.57 m³ for FALCON).
Execution Continuity: The system significantly reduced "infeasibility events" (optimizer failures) and premature terminations compared to baselines. In the Complex Office and Octa Maze, SuperFALCON+ achieved 0/10 premature terminations, whereas FALCON recorded 4/10 and 2/10 respectively. In the Duplex Office, SuperFALCON+ recorded 2/10 terminations, compared to 3/10 for both FALCON and EDEN.
Trade-offs: The improved resilience came at the cost of lower average flight speeds (1.05–1.27 m/s for SuperFALCON+ vs. 1.68–1.85 m/s for FALCON) and longer mission durations, reflecting a shift from rapid traversal to sustained, reliable exploration.
Ablation Studies: In the Octa Maze scenario, removing any of the three coordination components (Boundary Buffer, Watchdog Relay, or Target-Broker) resulted in 0/5 mission completions for the reduced configurations. In contrast, the full SuperFALCON+ system achieved 5/5 completions in the same ablation set, confirming that all three mechanisms are necessary for the observed resilience.
Real-World Validation: A physical quadrotor equipped with a Livox Mid-360 LiDAR and Jetson Orin NX successfully completed an outdoor exploration run in a campus environment with trees and foliage, maintaining operation despite trajectory oscillations near obstacles.

Significance and Claims

The paper claims that SuperFALCON+ demonstrates that resilience in autonomous exploration is not solely a function of better global pathfinding or faster local optimization, but rather the result of a well-defined coordination interface that manages the transfer of intent under local constraints.

The authors state that the system does not replace the exploration strategy of FALCON or the trajectory planner of SUPER. Instead, it coordinates their interaction to ensure that locally difficult requests are less likely to interrupt the exploration mission. The significance lies in the ability to sustain mapped-space growth and avoid premature termination in highly constrained environments where independent planners fail. The work suggests that future systems should treat the global-local interface as a critical design element, potentially adapting parameters based on measured progress and planner feedback.