Trajectory Guidance via Neural Networks
Overview
The effectiveness of mobile robots in realistic deployments is
dependent upon their ability to negotiate diverse environments in
which obstacles are dynamic, trajectories are occluded, and
lighting conditions may vary. Many existing control methods for
robot trajectory guidance utilize knowledge-based approaches that
rely on structured environments. While effective in simplistic
scenarios, these methods are severely limited by their
surroundings. Moreover, they do not generalize for real-world use
due to the inherent uncertainty in states encountered during
operation. Neural networks have been successfully used to build
autonomous navigation systems that can detect lane markings and
segment the ground plane (e.g., ALVINN). They can be trained to
produce control outputs based on noisy images where information is
contained in disparate areas and trajectories are discontinuous.
Contributions
- We created a system for adapting a line following robot to noisy, dynamic, and non-binary environments using two novel neural network architectures