[paper notes] mobile robot autonomous navigation experimental platform based on deep learning

Abstract

The work of this paper ：

1. Built a mobile robot experimental platform ;
2. An autonomous navigation method based on deep learning is designed $\to$ Input RGB Images , Directly output the control signal , Avoid complex feature engineering and planning strategies ;

key word

1. Mobile robots ;
2. Autonomous Navigation ;
3. Deep learning ;
4. Convolutional neural networks

Usually, mobile robot systems sense the surrounding environment through sensors , Such as laser sensor 、 ultrasonic sensor 、 Vision Sensors, etc , However, due to the limited carrying sensors , Most systems have insufficient decision-making and control capabilities .

The hardware mainly includes ： Raspberry pie 4B Upper computer 、STM32F103RC Lower machine 、 Motor and its drive 、 sensor

1 Mobile robot platform

1.1 Platform hardware system

project	Content
sensor （ Encoder ）	Hall encoder with incremental output
sensor （ camera ）	LETMC-520 camera
chassis	Mcnamm wheel , The hub shaft is formed with the roller rotating shaft $45°$ horn , It can ensure the omni-directional movement of the robot
The motor	GB37520 DC motor
Motor drive	TB6612FNG
Power Supply	Aircraft model battery
Upper computer	Raspberry pie 4B $\to$ Algorithm implementation runs 、 Collect sensor information 、 Lower computer communication
signal communication	USB2.0 two-way 、 Data frame format 、USB-TTL modular
Lower machine	STM32F103RC $\to$ Data collection 、 Chassis Control 、 signal communication

1.2 Platform software system

project	Content
OS	Ubuntu Mate 18.04
Chassis control node	subscribe /cmd_vel topic of conversation $\to$ Extract chassis target Linear velocity and angular velocity Information $\to$ Inverse kinematics $\to$ Motor's Target speed $\to$ Lower machine
Chassis control node	Get chassis motion state data , and Release Corresponding topic
Lower machine	Ｃ Language programming 、FreeRTOS Schedule tasks
Motor control algorithm	PID

2 Autonomous navigation method based on deep learning

2.1 Deep learning model

The four are $(128,64,64,16)$ The full connection layer of
Each floor also adopts ReLU Function to activate

Definition ：

1. $s(k)$$\to$ Steering control signal
2. $i(k)$$\to$ Model input RGB Images
3. $i(k)=f(s(k))$ The trained model is used $f$ Express
4. ${\omega }^{\ast }$$\to$ Target angular velocity of mobile robot
5. The inverse normalized signal is processed by first-order low-pass filtering to obtain the target angular velocity of the mobile robot ：
   $${\omega }^{\ast }=\alpha \beta s(k)+(1-\alpha ){\omega }^{\ast }(k-1)$$
   $\alpha =0.9$ And $\beta =1.4$

2.2 model training

Control robots $\to$ Observe - The action is right Data sets

project	Content
package	Keras、Tensorflow
normalization （ Images ）	yes ,$[0,1]$
normalization （ Steering control signal ）	yes ,$[-1,1]$
Data ratio （ corner ）	0.1494
Data ratio （ Straight Line ）	0.6304
Data ratio （ obstacle ）	0.2202
Dataset size	106623
Data set partition ratio	3：1：1
Data set content	80% $\to$ No obstacles + noise ;20% $\to$ Avoid obstacles + No noise is injected
Training process monitoring indicators	Mean square error MSE、 Mean absolute error MAE
Training loss function	Root mean square
Train the optimizer	RMSProp
Steering control signal threshold	0.1

The mean square error and mean absolute error on the test set are 0.039 and 0.102

3 Autonomous Navigation experiment

1. Autonomous Navigation node Pre use Keras、Tensorflow Read the trained model
2. subscribe /camera/rgb/image_raw Topic to Fixed period Get the of the camera RGB Images
3. RGB Image input model , Output steering control signal
4. Anti normalization And First order low-pass filter The target angular velocity of the mobile robot is obtained by processing ; Release **/cmd_vel** topic of conversation
5. ROS Chassis control node subscription in **/cmd_vel** Topic to control the motion of mobile robots .

Generalization experiment ： Avoid sudden obstacles $\to$ Empty bucket bottle

4 Conclusion