研究目的
To propose a novel Image-Based Visual Servoing (IBVS) controller for multirotor aerial robots using Deep Deterministic Policy Gradients (DDPG) to address limitations in convergence, stability, and gain design of classic approaches, and to validate it in simulated and real flight scenarios for object following applications.
研究成果
The proposed RL-IBVS controller demonstrates good capabilities in simulated and real flight scenarios, with a quick and direct transition from simulation to real flights. It shows versatility by working with different quadrotor models. Future work includes extending control to attitude commands and exploring other deep reinforcement learning algorithms.
研究不足
The reward function design is crucial and difficult, with success highly dependent on it. Training requires image stabilization, and direct use of actual image detections without stabilization led to unsuccessful policies. The approach is tested in indoor scenarios and may have limitations in outdoor or more dynamic environments.
1：Experimental Design and Method Selection:
The methodology involves designing a reinforcement learning agent based on DDPG to map image errors to velocity commands. The agent is trained in a Gazebo-based simulation environment using the RotorS simulator.
2：Sample Selection and Data Sources:
Training uses a simulated scenario with an AscTec Hummingbird quadrotor performing visual servoing over a cylindrical object. Real flight tests use a Parrot Bebop2 quadrotor.
3：List of Experimental Equipment and Materials:
Includes AscTec Hummingbird and Firefly quadrotors, Parrot Bebop2, Nvidia GeForce GTX 970 GPU, Intel Core i7-6700HQ CPU, OptiTrack motion capture system, and cameras with specified resolutions.
4：Experimental Procedures and Operational Workflow:
The agent is trained episodically with random initial positions, running at 20Hz, and actions are executed with added noise. Validation involves leader-follower experiments in simulation and real flights, comparing with classic and partitioned IBVS controllers.
5：Data Analysis Methods:
Performance is evaluated based on mean and standard deviation of image errors (ex, ey, eΦ), and results are compared across different IBVS approaches.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容-
Nvidia GeForce GTX 970
GTX 970
Nvidia
Used for training the RL-IBVS agent with TensorFlow.
-
AscTec Hummingbird
Hummingbird
AscTec
Used as the aerial robot platform for training the RL-IBVS agent in simulation.
-
AscTec Firefly
Firefly
AscTec
Used as the leader quadrotor in simulation experiments.
-
Parrot Bebop2
Bebop2
Parrot
Used as the aerial robot platform for real flight experiments.
-
Intel Core i7-6700HQ
i7-6700HQ
Intel
Used as the CPU in the ground computer for real flight experiments.
-
OptiTrack
OptiTrack
Used for gathering ground truth data in real flight experiments.
-
RotorS Gazebo
RotorS
Used as the simulation environment for training and testing the RL-IBVS agent.
-
TensorFlow
TensorFlow
Used as the deep learning library for implementing the DDPG algorithm.
-
ROS
ROS
Used as the middleware for communication between the agent and environment.
-
Aerostack
Aerostack
Used as the framework for integrating the RL-IBVS architecture with additional software components.
-
Parrot S.L.A.M. dunk
S.L.A.M. dunk
Parrot
Used for obtaining velocities estimation in real flight experiments.
-
登录查看剩余9件设备及参数对照表
查看全部
