Seeking Advice on Building a Custom Planner for Vision-Only Social Navigation Using Navigation GAN

Navigation Stack
1

Hello ROS community,

I am currently working on my master’s project, which focuses on vision-only social navigation. The goal is to enable a robot to navigate in dynamic environments by predicting the paths of pedestrians and planning its path toward a target using a Navigation GAN.

Here are some key points about my approach:

  • Pedestrian Path Prediction: I plan to use a Navigation GAN to predict the trajectories of detected pedestrians.
  • Custom Planner: My aim is to develop a custom planner that integrates these predictions to guide the robot.
  • Prediction Output: The model outputs the robot’s poses over 12 timesteps (equivalent to 8 seconds) into the future. Notably, the last pose does not necessarily represent the final goal.

I have a couple of questions and would greatly appreciate your insights:

  1. Dynamic Local Cost Map: Is it feasible to represent the detected pedestrians as a dynamic local cost map that updates during runtime? How would this approach compare to other methods for incorporating dynamic obstacles?
  2. Custom Planner Viability: Given that the Navigation GAN provides a series of future poses rather than a single goal pose, would implementing a custom planner be the most effective strategy? Are there existing tools or frameworks within ROS that could simplify this process, or should I focus on building a bespoke solution?

Any advice, recommendations, or relevant experiences would be highly valuable. Thank you for your time and assistance!

Best regards,Immanuel