Distributed Invariant Kalman Filter for Object-level Multi-robot Pose SLAM

A distributed left-invariant Kalman filter approach that utilizes semantic object-level 6-DoF estimations from Pose-CNN outputs and relative pose measurements between robots as filter observations. The method addresses uncertainty estimation issues caused by unknown correlations in inter-robot pose estimation through covariance intersection, ensuring system robustness when partial robots experience observation degradation.

We also proved the stability of the entire system: for each robot in the system, the expected square of its estimation error is bounded.

My Contributions: Extended invariant state dynamics to distributed Kalman filtering, conducted experimental validation, and developed the simulation environment.