Collision-free minimum snap optimized trajectory through a NeRF
Neural Radiance Fields (NeRFs) are a powerful 3D representation technique that leverages deep learning to reconstruct 3D scenes from a collection of 2D images. NeRFs encode the scene's geometry and appearance into a continuous function, allowing for the rendering of photorealistic images from arbitrary viewpoints. As a byproduct, the densities (translucencies) of the scene's objects at any point in the 3D environment are implicitly captured in the network. In this example, we demonstrate that such state-of-the-art learned models from computer vision research can be easily incorporated into optimization procedures using \lc. To showcase this, we present the problem of finding a collision-free trajectory through the densities represented by a learned NeRF, where densities below a predefined threshold are deemed as unobstructed regions within the environment.
Specify a problem configuration (start- and goal points) by changing the CASE variable in
nerf_trajectory_optimization.py.
Then run the optimization with
python nerf_trajectory_optimization.py
Three cases of a collision-free minimum snap optimized trajectory through a NeRF