TL;DR
This paper introduces an automated tethered perching system for aerial robots, inspired by birds. By combining analytical trajectory generation with reinforcement learning, the drone can safely wrap its tether around a branch and perform an energy-saving 180-degree flip to disarm, significantly extending operational time for data collection.
What Problem Does This Paper Address?
Long-term aerial data collection, particularly in challenging environments like forests, is severely limited by drone battery life. While drones can operate for short durations, extending their mission time for continuous monitoring requires innovative solutions. This paper tackles the challenge of enabling drones to conserve energy by perching on structures, mimicking avian behavior, and specifically addresses the complex maneuver of tethered perching on a branch. This involves not just landing, but also securely wrapping a tether and performing a controlled flip to disarm, all while avoiding collisions and minimizing energy expenditure.
Methods
The authors propose a two-phase approach for tethered perching:
- Approaching and Wrapping: This phase uses analytical methods to generate feasible trajectories. The core idea is to ensure sufficient tether contact with the branch for secure wrapping. The approach considers energy conservation and the physics of the tether, including its weight and the branch diameter. An equation is derived to relate tether length, drone velocity, and contact angle, aiming to minimize energy. A specific length requirement for wrapping is also calculated.
- Flipping Over: This phase is tackled using a reinforcement learning (RL) algorithm, specifically a soft actor-critic approach. The drone's state is defined by its relative position to the branch, roll angle, and velocity. The RL agent learns a policy to control the drone's roll angle to achieve a safe 180-degree flip and disarm. Two reward functions are employed: one for imitating a baseline solution and another for encouraging exploration and improving performance. The weights of these rewards are adjusted during training to balance imitation and exploration.
Datasets
Not discussed in the paper. The experiments were conducted in a controlled flight arena.
Key Insights and Ideas
- Bird-inspired energy saving: Mimicking how birds perch to conserve energy is a promising avenue for extending drone operational time, especially for long-duration data collection.
- Tethered perching as a viable strategy: Securely wrapping a tether around a branch and performing a controlled flip is a complex but achievable maneuver for autonomous drones.
- Hybrid analytical and RL approach: Combining analytical methods for initial trajectory planning with RL for fine-tuning and complex maneuvers (like the flip) offers a robust solution.
- Reinforcement learning for dynamic maneuvers: RL is well-suited for learning complex, dynamic control policies like the 180-degree flip, especially when dealing with uncertainties and the need for energy efficiency.
- Reward shaping for effective RL training: Using a combination of imitation and exploration rewards, with a gradual shift in their weighting, can lead to faster and more effective learning in RL agents.
Key Results
- The proposed algorithm successfully enabled the drone to perch on a branch in 16 out of 21 attempts in real-world experiments.
- The tethered perching maneuver, while requiring additional energy for the approach and wrapping, resulted in a net energy saving due to the ability to disarm and conserve power while perched. The paper states that the overall energy consumption during the tethering maneuver did not exhibit a significant difference compared to hovering.
- The RL-trained trajectories showed improved performance over a baseline manual approach, with the drone's angle increasing during the flip, leading to earlier attainment of final conditions and potential energy savings.
- The system demonstrated autonomous operation throughout the entire tethering maneuver.
Limitations and Future Work
- Tether collision: A significant cause of failure was the tether colliding with itself during the wrapping phase, particularly in the in-plane configuration.
- Non-slip condition: The assumption of a non-slip condition between the tether and the branch might not always hold true, potentially affecting energy calculations and requiring adjustments.
- Real-world complexity: The current experiments were conducted indoors. Future work should involve testing the proposed maneuver under real-life conditions, considering varying environmental factors, space, and take-off constraints.
- Recovering from upside-down position: The paper suggests exploring mechanisms for recovering from the upside-down perching position, including cutting the tether to place the sensor on the environment.
- Varying parameters: Future research could involve varying rope length and thrust during training to better simulate real-world conditions and improve the algorithm's adaptability.
- Geometry modification: Modifying the geometry of the weight or employing figure-of-eight trajectories could mitigate the risk of tether self-collision.
Summary updated