This is intended to be a lightweight 2D slam framework for educational purposes. Some experiments can be found in visualization, but the actual SLAM sim is in PySLAM/world_sim.py
to run call python3 world_sim.py
click on the map to place a new robot waypoint. True position in green, red is dead reckoning, and blue is graphslam output.
requires opencv, pyglet, and numpy
This is a very simple framework for fun than anything else. It uses a graphslam approach without weights. Features are based on sensor spikes.
The code is extremely unoptimized, but can still process hundreds of landmarks and positions in a few 100ms.
(note that the right image is inverted along the y axis)
I'm probably going to leave this project as is. There isn't much reason to add much to it as SLAM is a very well "solved" problem with considerably better solutions than this one. However, if improvements were to be made, the following would help:
- Overall refactoring of code to help clarity and organization
- Use of sparse matrices would allow dynamic additions instead of rebuilding the matrix every time
- Better feature identification
- Weighted measurements to rely more heavily on waypoints than deltas
- Graphical visualization cleaned up and separated into a separate module
- Use of KD-tree or quadtree (or just gridmap?!) to store landmarks to speed up landmark matching (likely a very large bottleneck)