Deep Learning applied to cooperative motion planning in environments shared with humans

Autonomous navigation has become, in the last years, one of the main field of research for both transportation and robotic purposes, employing all kind of terrain, maritime and aerial vehicles. An example of its growing importance is the recent development of autonomous driving solutions that is been carried out by most of the world leader car makers.A common aspect of such applications is the need of involving cooperation with other agents, that might be robots or humans, in order to accomplish the given task. Therefore, having at our disposal effective models that accurately predicts other agents actions becomes crucial in this context.The aim of this work is the design of a cooperative motion planning algorithm that comprises a series of required features in order to be employed in an autonomous navigation scheme. Among these features, we will focus on the ability to reason about neighboring agents and static obstacles, including observable information from other agents and characteristic data of the ego agent, and performing multi-hypothesis estimates.In this work, we have chosen the task of navigating through human crowds in complex environments. Therefore, our goal consists in designing a cooperative motion planning algorithm capable of modeling human motion behavior, including the earlier mentioned characteristics.The developed approach is an extension of a current state-of-the-art method designed by Alexandre Alahi, known as Social LSTM.This algorithm is composed by an artificial neural network that captures cooperative motion behavior by sharing information of the neighboring agents latent representations, and returns unimodal stochastic estimations of the future positions of the agents. This baseline method is going to be extended in order to include the missing mentioned qualities that are desired in a path prediction model within the context of autonomous navigation. These extensions are implemented through simulations of crowded environments in complex scenarios by using the open source crowd simulator Pedsim. The generated datasets, together with some real world recorded scenarios from public datasets, allowed us to test the presented algorithms in order to tune the corresponding hyperparameters of the neural network and design the different employed network architectures. Due to the lack of an official code implementation of the Social LSTM from Alexandre Alahi, it has been considering, as starting point, the Social LSTM implementation publicly available from Anirudh Vemula.In the presented results can be observed how the mentioned characteristics included in the path prediction model increases the inference accuracy of future trajectory distributions, evaluated using the obtained log-likelihood. This increment of the performance is translated into a 45% increase of the log-likelihood value in comparison with the Social LSTM baseline model in the evaluated datasets.These results indicate that the studied state-of-the-art path prediction methods, like the mentioned Social LSTM trained over datasets containing simple scenarios have a large margin of improvement when executed in more realistic situations.In following works, the implemented contributions together with some proposed future works might be employed to raise the performance of autonomous navigation algorithms in real-world environments.Mechanical Engineering | Systems and Contro