This work presents a case study of a learning-based approach for target driven map-less navigation. The underlying navigation model is an end-to-end neural network which is trained using a combination of expert demonstrations, imitation learning (IL) and reinforcement learning (RL). While RL and IL suffer from a large sample complexity and the distribution mismatch problem, respectively, we show that leveraging prior expert demonstrations for pre-training can reduce the training time to reach at least the same level of performance compared to plain RL by a factor of 5. We present a thorough evaluation of different combinations of expert demonstrations, different RL algorithms and reward functions, both in simulation and on a real robotic platform. Our results show that the final model outperforms both standalone approaches in the amount of successful navigation tasks. In addition, the RL reward function can be significantly simplified when using pre-training, e.g. by using a sparse reward only. The learned navigation policy is able to generalize to unseen and real-world environments.
Abstract-This paper reports on a data-driven motion planning approach for interaction-aware, socially-compliant robot navigation among human agents. Autonomous mobile robots navigating in workspaces shared with human agents require motion planning techniques providing seamless integration and smooth navigation in such. Smooth integration in mixed scenarios calls for two abilities of the robot: predicting actions of others and acting predictably for them. The former requirement requests trainable models of agent behaviors in order to accurately forecast their actions in the future, taking into account their reaction on the robot's decisions. A human-like navigation style of the robot facilitates other agents-most likely not aware of the underlying planning technique applied-to predict the robot motion vice versa, resulting in smoother joint navigation. The approach presented in this paper is based on a feature-based maximum entropy model and is able to guide a robot in an unstructured, real-world environment. The model is trained to predict joint behavior of heterogeneous groups of agents from onboard data of a mobile platform. We evaluate the benefit of interaction-aware motion planning in a realistic public setting with a total distance traveled of over 4 km. Interestingly the motion models learned from human-human interaction did not hold for robot-human interaction, due to the high attention and interest of pedestrians in testing basic braking functionality of the robot.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.