TL;DR
This paper presents a novel decentralized approach for micro aerial vehicles (MAVs) to simultaneously determine their positions and navigate in formation, using only onboard sensors like cameras, IMUs, and range sensors. This method allows teams of MAVs to coordinate complex maneuvers without relying on external infrastructure or centralized control, achieving centimeter-level accuracy.
What Problem Does This Paper Address?
Coordinating a team of robots, especially small, agile micro aerial vehicles (MAVs), for tasks like surveillance or search and rescue is challenging. Traditional methods often rely on a central "brain" to process all data and send commands, which doesn't scale well and is prone to single points of failure. Furthermore, relying on external positioning systems like GPS can be unreliable in indoor or obstructed environments. This paper tackles the problem of enabling MAVs to localize themselves and maintain formations autonomously, using only their onboard sensors and communicating with their immediate neighbors. This is crucial for robust, scalable, and flexible multi-robot operations.
Methods
The core of this work lies in a two-pronged, decentralized approach:
- Decentralized Cooperative Localization: Instead of a central system, each MAV estimates its own pose (position and orientation) by fusing data from its onboard Inertial Measurement Unit (IMU), Visual-Inertial Odometry (VIO), and range measurements to nearby MAVs. This is formulated as a factor graph optimization problem, solved using an asynchronous block coordinate descent (BCD) method. This means each MAV iteratively refines its estimate based on its own measurements and information from its neighbors, without needing strict synchronization or sharing all raw sensor data. A key innovation is the use of odometry preintegration to efficiently combine high-frequency VIO estimates into a single relative pose between time steps.
- Decentralized Formation Control: Once localized, each MAV generates control commands to maintain a desired formation. This is also framed as a factor graph inference problem. It incorporates a "motion prior" derived from a continuous-time Gaussian Process regression, which predicts the MAV's future state based on its current estimate and uncertainty. Crucially, this approach accounts for the uncertainty in the MAV's own state estimation and its neighbors' states. This probabilistic approach allows for smoother control and better performance, especially when sensor measurements are noisy or environments are dynamic. The formation is specified by desired inter-agent distances, and the control objective is to minimize the error between desired and actual distances.
Datasets
Not discussed in the paper. The authors conducted simulations and real-world experiments using their own MAVs and environments.
Key Insights and Ideas
- Decentralization is Key for Scalability and Robustness: The paper strongly advocates for decentralized solutions, showing that MAVs can achieve high performance without a central controller, making the system more resilient to individual failures and easier to scale to larger teams.
- Accounting for Uncertainty is Crucial for Formation Control: Simply trying to enforce rigid formation distances can lead to poor performance when sensor noise and estimation errors are present. By explicitly modeling and incorporating this uncertainty into the control objective, the system achieves more stable and accurate formations.
- Range Sensors are Powerful for Decentralized Coordination: Despite their simplicity, range sensors (like UWB) provide essential inter-agent measurements that, when combined with onboard VIO and IMU data, enable robust, infrastructure-free localization and formation control.
- Factor Graphs are a Versatile Framework: The paper effectively demonstrates the power of factor graphs for unifying localization and control problems, allowing for principled integration of various sensor modalities and probabilistic reasoning.
Key Results
- Localization Accuracy: The decentralized cooperative localization (dcl) method achieved positioning Root Mean Square Error (RMSE) comparable to centralized batch trajectory estimation, with errors in the order of decimeters (e.g., ~0.1m).
- Formation Accuracy: The proposed formation control approach resulted in lower formation errors (FRE) compared to a baseline gradient control method. In simulations, formation RMSE was around 0.05m, and in real experiments, it was around 0.1-0.2m depending on the trajectory.
- Computational Efficiency: The decentralized approach is computationally efficient, with compute times per agent showing minimal increase even with a larger number of neighbors or higher range update rates. This suggests good scalability.
- Real-World Performance: The approach was successfully demonstrated in diverse indoor and outdoor environments, including GPS-denied settings and areas with potential wind disturbances, executing formation flights with MAVs.
Limitations and Future Work
- Obstacle Avoidance: The current work does not explicitly address obstacle avoidance. The authors plan to incorporate this into the factor graph framework for future work.
- Speed Limitations: The reactive nature of the control, combined with communication and computation delays, leads to larger tracking errors at higher speeds.
- Wind Disturbances: While tested in environments with potential wind, the paper notes that wind disturbances can impact performance.
- Future Work: The authors aim to investigate increasing the speed of formation flights and incorporating obstacle avoidance. They also plan to explore more complex formation shapes and missions.
Summary updated