A Real-Time Game Theoretic Planner for Autonomous Two-Player Drone Racing

Riccardo Spica, Davide Falanga, Eric Cristofalo, Eduardo Montijano, Davide Scaramuzza, Mac Schwager

Introduction

Autonomous drone research has traditionally focused on time trials—optimizing a single vehicle's trajectory to navigate a course as rapidly as possible. However, competitive racing (such as "rotocross") introduces complex non-cooperative interactions between multiple agents. In these scenarios, speed is not the sole factor; drones must strategically react to opponents who may actively block paths or contest racing lines. This paper presents a Game Theoretic Planner (GTP) designed to handle these reciprocal dynamics in real-time without direct communication between vehicles.

The Challenge of Non-Cooperative Interaction

The primary difficulty in multi-drone racing is predicting an opponent's intent. Treating an opponent as a static obstacle leads to overly passive behavior, while assuming they will yield can result in collisions. The GTP addresses this "reciprocal dance" by modeling the interaction as a game where each player's strategy depends on their prediction of the other's response. The core challenge is to find a balance between aggressive maneuvering and safe collision avoidance.

Methodology

The proposed GTP operates at 20Hz and relies on two primary mechanisms: Iterated Best Response (IBR) and sensitivity analysis.

Iterated Best Response (IBR)

To avoid the computational burden of solving a full multi-player game, the system employs an IBR scheme. In this approach, players "take turns" in a simulated optimization loop. One player optimizes their trajectory assuming the other's path is fixed; subsequently, the second player optimizes their path relative to the first player's new trajectory. This process repeats until a Nash Equilibrium is reached—a state where neither player can improve their outcome by unilaterally changing their strategy.

Sensitivity Analysis and Strategic Play

A key innovation is the use of sensitivity analysis to account for how one player's constraints influence another. By approximating how the ego-drone's positioning forces the opponent to alter their path for collision avoidance, the GTP can execute aggressive maneuvers. For instance, a drone can "push" an opponent off-course or strategically block a passing lane by understanding the opponent's likely reactive constraints. This behavior is tuned via an "aggressiveness factor" ($\alpha$) in the cost function.

Experimental Validation

The system was evaluated through extensive simulations and real-world hardware experiments.

Simulation and Results

In 900 simulations using the RotorS quadrotor simulator, the GTP was compared against standard Model Predictive Control (MPC) and Reciprocal Velocity Obstacles (RVO). The GTP consistently outperformed these baselines. In scenarios where the GTP drone was slower but started ahead, it successfully blocked faster MPC-controlled drones in nearly all trials. Conversely, when faster, the GTP drone achieved successful overtakes in approximately 30% of cases where naive planners remained stuck behind the opponent.

Hardware Implementation

Real-world tests were conducted using DJI F330 quadrotors equipped with monocular cameras and onboard Odroid computers. The drones utilized vision-based tracking of colored markers and a Kalman Filter to estimate the opponent's 3D state, proving the system's feasibility on standard hardware.

Conclusion and Future Work

The GTP demonstrates that game-theoretic motion planning is feasible for high-speed, real-time robotics. By exploiting the reactivity of opponents through sensitivity analysis, autonomous drones can move beyond simple obstacle avoidance toward sophisticated strategic racing. While the current model assumes opponent similarity and can face local minima in non-convex optimization, future work aims to scale the system to $N$ players and incorporate online learning of opponent-specific cost functions.

Summary updated

BibTeX
@inproceedings{spica2018real,
  author = {Riccardo Spica and Davide Falanga and Eric Cristofalo and Eduardo Montijano and Davide Scaramuzza and Mac Schwager},
  title = {{A Real-Time Game Theoretic Planner for Autonomous Two-Player Drone Racing}},
  booktitle = {Robotics: Science and Systems},
  year = {2018},
  doi = {10.15607/rss.2018.xiv.040},
}
Figure