Author: Antoni Rosinol
- Robot able to answer high level queries
- Kimera = metric / semantic SLAM with stereo images and IMU
- goal geometric + semantic accuracy
- first step: sparse odometry to generates 3D mesh
- Uses CNN for semantic
- 3D semantic mesh doesn't the concept of object => it is not actionnable
- 3D dynamic scene graph:
- layer one: metric semantic mesh
- layer two: objects and agents (detects and tracks agent like humans/objects)
- layer three: places and structure (free spaces rps as topological map, ready to be used for path planning)
- layer four: rooms
- layer five: buildings
- How to optimize a 3D mesh? with pose graph optimization on each vertices
- Update 3D mesh is really hard
Futur of SLAM:
- learning in SLAM is loosely coupled => tightly coupling?
- idea compute gradient of weights wrt SLAM problem
- differentiable rendering for SLAM
Semantic = what a human can tel about of a pixel