Core Concepts
A neural network-based approach to learning socially-aware crowd navigation strategies for autonomous robots in real-world environments.
Abstract
The authors present a method for teaching autonomous mobile robots to successfully navigate human crowds using a neural network-based approach. The key highlights and insights are:
Crowd navigation requires more than just path planning and obstacle avoidance - it needs to account for social norms and human behavior, which can vary across contexts.
The authors use a Convolutional Neural Network (CNN) that takes a top-down image of the scene as input and outputs the next action for the robot in terms of speed and angle. This allows the robot to learn strategies specific to the context.
To capture real-world robot-human interactions, the authors collect training data by tele-operating the robot through various crowd scenarios in a university hallway, including before, during, and after class times.
The authors perform camera calibration, homographic reprojection, and human/robot detection to preprocess the data into a format suitable for training the CNN.
After extensive hyperparameter tuning, the final CNN architecture consists of 3 convolutional layers followed by 3 fully connected layers. The model is trained using Mean Squared Error loss.
The trained model is able to learn appropriate speed and rotation strategies for navigating the hallway, with an average deviation of 12 cm/s in speed and 7 degrees in rotation compared to the baseline.
Due to a mechanical issue with the robot, the authors were unable to complete the planned real-world evaluation of the trained model. However, they outline plans for future work, including closing the loop to autonomously control the robot, adding more sensors, and expanding to other environments.
Stats
The average baseline speed of the robot was 22.617376 cm/s, and the average baseline rotation was 13.952758 degrees.
Using the trained neural network, the average speed was 10.503133 cm/s, and the average rotation was 6.349384 degrees.