Project 3 Image 1

Preemptive planning and Dynamic obstacle avoidance with SGANs and MPC

Robot motion planning in dynamic environments is a very challenging problem that requires the​ robot to generate safe and efficient paths to follow. Where the safety of pedestrians or humans is​ associated with robot collisions, the factor of the robot’s efficiency plays a pivotal role in minimizing​ pause time, path navigation, and utilization of resources such as batteries.​ Keeping these parameters into consideration, and the example of the Starship food delivery robots,​ which tends to halt if people come in front of it, we have solved this problem by integrating SGANs​ (Social Generative Adversarial Networks) for human movement predictions and obstacle modeling,​ traditional motion planning algorithms and approaches such as RRT*/modified RRT* paired with​ spline interpolation, and motion control using Non-Linear Model Predictive Controller (NMPC) to​ dynamically avoid obstacles and reach the target location.​ Our approach and experimental results indicate that the chance of collision drastically drops to​ approximately 70% in an environment with up to 6 people.​

Project 3 Image 2

For the baseline, we utilized Social GAN: Socially Acceptable Trajectories with Generative Adversar-​ ial Networks by Agrim Gupta et al. [1], which assists to predict the motion behavior of pedestrians.​ The paper uses GANs to overcome the difficulties in approximating intractable probabilistic compu-​ tation and behavioral inference. The proposed GAN in the SGAN paper is an RNN Encoder-Decoder​ generator and an RNN-based encoder discriminator with a variety loss to encourage the GAN’s​ generative network to spread its distribution and cover the space of possible paths while remaining​ consistent with the observed inputs and a new pooling mechanism which learns a "global" pooling​ vector that encodes subtle cues for all participants in a scene. This model uses accuracy, speed, and​ the ability of the model as metrics to generate a variety of socially acceptable trajectories based on​ experiments on several publicly available real-world crowd datasets​

View on GitHub