Vision-Less Sensing for Autonomous Micro-Drones

Simon Pikalov, Elisha Azaria, Shaya Sonnenberg, Boaz Ben-Moshe, Amos Azaria

TL;DR

This paper introduces BAT, a "Blind Autonomous Tiny-drone" that navigates and maps indoor environments using only LiDAR and optical flow sensors, eschewing cameras entirely. This bio-inspired approach allows for fully autonomous operation even in complete darkness, demonstrating a novel path towards more robust and versatile micro-drone applications.

What Problem Does This Paper Address?

Current autonomous drones, especially micro-drones, often rely heavily on cameras for navigation and obstacle avoidance. While effective in many scenarios, camera-based systems can struggle in low-light conditions, dynamic environments, or when visual data is obscured. Furthermore, the computational demands of processing high-resolution visual data can be a bottleneck for resource-constrained micro-drones. The authors aim to address these limitations by developing a vision-less micro-drone inspired by the navigation strategies of flying animals, which can operate effectively without relying on visual input. This opens up possibilities for drones to perform complex tasks in environments where cameras are impractical or insufficient.

Methods

The core of the BAT system is its vision-less sensing and control architecture. Instead of cameras, BAT utilizes a combination of sensors:

  • Multi-Ranger LiDAR Deck: A custom PCB with five VL53L1X Time-of-Flight (ToF) LiDAR sensors strategically placed to provide 360-degree range coverage (front, back, left, right, and up). These sensors measure distances up to 4 meters at 25 Hz.
  • Onboard Optical Flow Sensor: Integrated into the base Tello drone, this sensor measures the drone's relative velocity by analyzing consecutive frames. While not directly accessible for navigation, it aids in maintaining stable flight.
  • Time-of-Flight (ToF) Range Sensor: Also part of the base Tello, this sensor measures the drone's height from the ground, crucial for avoiding obstacles below and navigating vertical changes.
  • Barometer: Used to determine relative altitude with respect to the takeoff point.

These sensors feed into an ESP32 microcontroller, which runs the autonomous control algorithms. The control strategy is based on a wall-follower algorithm, adapted to select states like "Rotate C.C.W.", "Emergency", "Tunnel", "Turn C.W.", and "Fly Forward" to navigate and explore. PID controllers are used for Pitch, Roll, and Throttle, while Yaw is managed by the state machine. For mapping, the collected sensor data is transmitted to a mapping station (via LoRa communication) to construct a 2D and 3D geometric model of the environment.

Datasets

Not discussed in the paper. The experiments were conducted in a custom-built 3D simulation environment (Microsoft AirSim) and in real-world indoor settings across seven different buildings.

Key Insights and Ideas

  • Vision-less navigation is feasible for complex tasks: The paper demonstrates that micro-drones can achieve sophisticated navigation, obstacle avoidance, and mapping without cameras, challenging the conventional reliance on visual sensors.
  • Bio-inspiration as a design principle: Mimicking the sensory and navigation strategies of flying animals (like bats) provides a robust framework for developing autonomous systems that can operate in challenging, unstructured environments.
  • Sensor fusion for robust perception: Combining LiDAR, optical flow, and barometer data creates a comprehensive environmental perception that compensates for the limitations of individual sensors, especially in low-light conditions.
  • Efficient computation on microcontrollers: The use of an ESP32 microcontroller highlights the possibility of performing complex autonomous tasks on low-power, embedded hardware, making the system more accessible and cost-effective.
  • Mapping without visual landmarks: The ability to generate detailed 2D and 3D maps using only range data is a significant step towards autonomous exploration in environments where visual features are scarce or unreliable.

Key Results

  • The BAT successfully explored and mapped complex indoor environments, including buildings and tunnels, in both simulation and real-world tests.
  • In simulation, BAT explored a 600 m² building over 5 minutes without collisions.
  • In real-world tests, BAT flew for approximately 5 minutes (limited by battery life) without crashing and without returning to the same spot twice.
  • The system demonstrated successful navigation up stairs in a two-story building.
  • The generated maps, both 2D and 3D, provided a geometric representation of the explored environments, with an expected error of 1-2 meters for maps generated after 5 minutes of flight.

Limitations and Future Work

Limitations:

  • Limited sensor range: The LiDAR sensors have a maximum range of 4 meters, which might limit exploration in very large open spaces.
  • IMU drift: The gyro measuring yaw has an expected drift of about 1°/min, which can affect long-term localization accuracy.
  • Optical Flow limitations: The optical flow sensor's performance is correlated to light and ground texture conditions, and its error is below 10% of the distance.
  • Barometer drift: The barometer can have a drift of up to 10 cm per minute, though the expected attitude error is usually smaller.
  • Mapping accuracy: The output map has an expected error of 1-2 meters after 5 minutes of flight.

Future Work:

  • Graph-SLAM integration: Developing a Graph-SLAM algorithm on the sparse LiDAR data to improve the drone's ability to learn and recognize known places and navigate accordingly.
  • Scaling down: Further reducing the size of BAT to enhance maneuverability and crash resistance.
  • Advanced sensors: Replacing LiDARs with ToF matrix ranging sensors (e.g., Pico Flexx) for more precise 3D reconstruction.
  • Multi-drone cooperation: Extending the system to multiple BATs for accelerated mapping and cooperative missions, including data sharing.
  • Reinforcement Learning: Utilizing reinforcement learning to improve the system's robustness and performance in dynamic environments and complex objective functions, potentially leading to more adaptive autonomous behavior.

Summary updated

Abstract

This work presents a concept of intelligent vision-less micro-drones, which are motivated by flying animals such as insects, birds, and bats. The presented micro-drone (named BAT: Blind Autonomous Tiny-drone) can perform bio-inspired complex tasks without the use of cameras. The BAT uses LIDARs and self-emitted optical-flow in order to perform obstacle avoiding and maze-solving. The controlling algorithms were implemented on an onboard micro-controller, allowing the BAT to be fully autonomous. We further present a method for using the information collected by the drone to generate a detailed mapping of the environment. A complete model of the BAT was implemented and tested using several scenarios both in simulation and field experiments, in which it was able to explore and map complex building autonomously even in total darkness.
BibTeX
@inproceedings{pikalov2021vision,
  author = {Simon Pikalov and Elisha Azaria and Shaya Sonnenberg and Boaz Ben-Moshe and Amos Azaria},
  title = {{Vision-Less Sensing for Autonomous Micro-Drones}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=lLYGdyDs29c},
}
Figure