⚡ electrical engineering

Nonlinear receding-horizon differential game for drone racing along a three-dimensional path

This paper proposes a nonlinear receding-horizon differential game (NRHDG) framework that integrates adversarial prediction, a novel path-following formulation, and dynamic maneuver switching to significantly outperform traditional nonlinear model predictive control in competitive three-dimensional drone racing scenarios.

Original authors: Kijin Sung, Kenta Hoshino, Akihiko Honda, Takeya Shima, Toshiyuki Ohtsuka

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kijin Sung, Kenta Hoshino, Akihiko Honda, Takeya Shima, Toshiyuki Ohtsuka

Original paper licensed under CC BY 4.0 (http://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 high-speed drone race through a complex, 3D obstacle course. The goal isn't just to fly fast; it's to outsmart a rival drone that is trying to block you, while you are trying to pass them. This paper presents a new "brain" for racing drones that is better at this mental chess game than previous methods.

Here is the breakdown of their solution using simple analogies:

1. The Problem: The "Stiff" vs. The "Smart"

Previous drone controllers were like stiff, single-minded runners. They were great at following a specific track perfectly, but if another runner tried to cut them off, they didn't know how to react. They assumed the other runner would just keep running in a straight line, which is rarely true in a real race.

The authors wanted a controller that acts like a competitive sprinter. This sprinter needs to:

Stay on the track (but not so rigidly that they can't dodge).
Predict what the rival will do next.
Decide in real-time: "Should I try to pass them now, or should I block them from passing me?"

2. The Solution: The "Game-Theory" Brain

The authors created a system called NRHDG (Nonlinear Receding-Horizon Differential Game). Think of this as a simulator inside the drone's head that runs a "what-if" game every millisecond.

The Worst-Case Scenario: Instead of guessing what the opponent will do, the drone assumes the opponent is a master strategist trying to ruin its race. It asks, "If my opponent tries their absolute best to stop me, what is the best move I can make to win anyway?"
The Saddle Point: In game theory, there is a "saddle point" where both players have found their best possible move. The drone calculates this instantly. If the opponent changes their strategy, the drone is already prepared for it.

3. Three Key Innovations

A. The "Ghost Marker" (Path Following)

Usually, to stay on a 3D track, a drone has to constantly calculate the shortest distance to the line, which is like trying to measure the distance from a moving car to a winding road while driving at 100 mph. It's slow and computationally heavy.

The authors invented a "Ghost Marker" system. Instead of measuring the distance, they attach a virtual marker to the track that moves along with the drone. The drone simply follows this marker.

Analogy: Imagine you are walking on a moving walkway at the airport. Instead of constantly checking how far you are from the center line, you just keep your eyes on a specific spot on the floor that moves with you. This makes the math much faster and smoother, allowing the drone to react instantly.

B. The "Shape-Shifting" Strategy (The Potential Function)

The drone needs to switch between two modes: Overtaking (offense) and Obstructing (defense).

The Old Way: Drones were hard-coded. If you were behind, you tried to pass. If you were ahead, you just kept flying.
The New Way: The authors created a "magnetic field" (a potential function) around the rival drone.
- When you are behind: The field pushes you away from the rival (like a magnet repelling) so you can find a gap to pass.
- When you are ahead: The field pulls you toward the rival's path (like a magnet attracting) so you can block their lane.
- The Magic: This switch happens smoothly and automatically based on where you are relative to the opponent, without the drone needing to "think" about which mode to switch to.

C. The "Fair Race" Scorecard

To prove their new brain was better, they needed a fair way to compare it against the old "stiff" brain (NMPC).

The Flaw in Old Tests: Previous studies often just swapped controllers between two drones (Drone A uses Brain X, Drone B uses Brain Y) and saw who won. This was unfair because one drone might have started with a better position.
The New Test: They ran four specific scenarios:
1. New Brain vs. New Brain
2. New Brain vs. Old Brain
3. Old Brain vs. New Brain
4. Old Brain vs. Old Brain
  By keeping the starting positions identical and only changing the "brain," they could prove that the New Brain was the reason for the better performance, not luck.

4. The Results: The Proof

They ran 400 simulated races with random starting positions and different speed advantages.

The Outcome: The "New Brain" (NRHDG) consistently won.
Offense: When trying to pass, the New Brain got further ahead than the Old Brain before completing the pass.
Defense: When trying to block, the New Brain successfully slowed down the opponent more effectively than the Old Brain.
Statistical Significance: The results weren't just "sometimes better." The data showed a clear, statistically significant advantage. Even when the "New Brain" drone was slightly slower, it still managed to outsmart the opponent more often than the "Old Brain" did.

Summary

This paper introduces a drone racing controller that treats the race like a dynamic game of chess rather than a simple driving task. By using a "Ghost Marker" to follow the track efficiently and a "Shape-Shifting" strategy to switch between attacking and defending, the drone can outmaneuver opponents who are using older, less adaptable software. The simulations prove that this "game-theoretic" approach is statistically superior in both attacking and defending roles.

Technical Summary: Nonlinear Receding-Horizon Differential Game for Drone Racing

Problem Statement

Drone racing presents significant control engineering challenges due to the requirement for high-speed navigation through complex three-dimensional paths under nonlinear dynamics. While existing methods, including trajectory optimization and learning-based approaches, have achieved precise single-drone navigation, they largely fail to address multi-agent competitive interactions. Specifically, current frameworks lack a feedback control mechanism that simultaneously handles nonlinear drone dynamics and adversarial behaviors such as overtaking and obstructing. Furthermore, conventional path-following methods often rely on computationally expensive iterative searches for orthogonal projections or are limited to two-dimensional paths, while existing differential game approaches for racing often assume fixed roles (overtaking or obstructing) and rely on simplified kinematic models or explicit path features that increase computational costs.

Methodology

The study proposes a Nonlinear Receding-Horizon Differential Game (NRHDG) framework designed to address these limitations. The methodology integrates three core components:

1. Unified Path-Following Formulation

The authors derive a novel dynamical model for the orthogonal projection of a drone's position onto a three-dimensional path. Unlike traditional methods that require iterative minimization of distance during online control, this approach treats the path parameter ( $\theta_d$ ) and the associated arc length ( $\sigma$ ) as state variables.

Projection-Point Dynamics: By differentiating the stationary condition for the projection point, the authors derive ordinary differential equations (Theorem 1) that dynamically update the projection point and arc length.
Augmented State: These dynamics are embedded directly into the drone's state equation, creating a 15-dimensional augmented state vector. This eliminates the need for separate projection optimization or complex coordinate transformations (e.g., Frenet-Serret frames) during the control loop.

2. Dynamic Role Switching via Potential Functions

To enable flexible competitive strategies, the study introduces a smooth potential function ( $G_O$ ) that allows drones to switch between overtaking and obstructing modes based on their relative positions.

Mechanism: The function depends on the difference in path parameters ( $\theta_\Delta$ ) and the norm of the difference in deviations from the path ( $R$ ).
Behavior: When the ego drone is behind ( $\theta_\Delta$ is large), the potential encourages avoidance and overtaking (maximizing $R$ ). When the ego drone is ahead ( $\theta_\Delta$ is small), the potential encourages obstructing (minimizing $R$ ).
Advantage: The smoothness of this function makes it suitable for sensitivity-based numerical optimization without introducing hard collision-avoidance constraints, which are computationally burdensome.

3. Control Frameworks: NMPC vs. NRHDG

The study compares two control strategies:

Nonlinear Model Predictive Control (NMPC): The ego drone optimizes its trajectory assuming the opponent follows a simplified, fixed prediction model (constant path-parameter rate and constant distance from the path).
NRHDG: The problem is formulated as a two-player, zero-sum differential game. The ego drone minimizes a cost function while explicitly assuming the opponent maximizes the same cost function. Both drones are modeled using the full nonlinear augmented state equations. The solution seeks a saddle-point strategy, allowing the ego drone to counter the opponent's worst-case actions in real time.

4. Performance Metrics

New metrics are introduced to evaluate overtaking and obstructing performance independently:

Overtaking Performance: Measured by the arc-length progress of the rear-start drone at the moment it overtakes the front-start drone.
Obstructing Performance: Measured by the arc-length progress of the rear-start drone when being obstructed by the front-start drone.
Comparison Protocol: The study compares controllers by running races where the same front-start controller faces different rear-start controllers (and vice versa) under identical initial conditions, isolating the controller's performance from the opponent's identity.

Key Contributions

Unified Dynamical Model: A computationally efficient path-following model based on projection-point dynamics that embeds directly into the state equation, avoiding iterative distance minimization and extending to arbitrary smooth curves.
Smooth Potential Function: A custom potential function enabling real-time switching between overtaking and obstructing strategies, tailored for sensitivity-based optimization of fully nonlinear models.
Systematic Performance Metrics: A novel evaluation framework that attributes performance differences solely to the controllers by holding initial conditions and opponent identities constant across comparative races.

Results

Numerical simulations were conducted on a 3D path using a Parrot MamboFly model. The NRHDG and NMPC controllers were tested under various initial conditions and speed advantages (controlled via input weights).

Overtaking Performance: NRHDG consistently outperformed NMPC. In races where NRHDG was the rear-start drone, it achieved significantly greater arc-length progress before overtaking the opponent compared to NMPC. Statistical analysis of 100 random initial conditions showed that the 95% confidence intervals for the mean performance difference excluded zero, indicating a statistically significant advantage.
Obstructing Performance: NRHDG demonstrated superior obstructing capabilities. When acting as the front-start drone, NRHDG successfully limited the progress of the rear-start opponent more effectively than NMPC did. The 95% confidence intervals for obstructing performance differences were consistently negative, confirming that NRHDG obstructs opponents more effectively.
Robustness: The superiority of NRHDG held across different levels of speed advantage for the rear-start drone, though the magnitude of the advantage varied. NRHDG was particularly effective when the rear-start drone had a smaller speed advantage, leveraging the worst-case modeling to maximize progress.

Significance and Claims

The paper claims that the proposed NRHDG framework effectively captures competitive interactions in drone racing, offering a statistically significant performance improvement over standard NMPC in both offensive (overtaking) and defensive (obstructing) scenarios.

The study emphasizes that explicitly modeling adversarial behavior through a differential game approach allows the controller to optimize against worst-case opponent actions, a capability lacking in standard MPC which relies on fixed opponent models. The results suggest that for competitive multi-agent systems, explicitly accounting for adversarial dynamics is crucial for achieving optimal performance. The authors note that while the current study focuses on two drones on a single path, the developed principles have potential applications in broader domains involving dynamic multi-agent interactions, such as autonomous vehicle coordination and air traffic management, though these are presented as potential future directions rather than current experimental validations.