Proprioceptive Invariant Robot State Estimation
7 November 2023
Presents DRIFT, a real-time proprioceptive state estimation framework
Uses invariant Kalman filtering for improved consistency
Includes tutorial introduction to invariant filtering
Develops dead reckoning modules for diverse robot types
Adds contact estimation and gyro filtering for low-cost robots
Shows accurate long-term tracking without perception
Proprioceptive Robot State Estimation
This paper presents DRIFT, a real-time proprioceptive robot state estimation framework. DRIFT uses only data from onboard sensors like IMUs and joint encoders to track a robot's orientation, velocity, and position. The method is based on invariant Kalman filtering, which exploits symmetries in the state space for improved consistency. The authors provide a tutorial introduction to invariant filtering, making the approach more accessible. They develop DRIFT modules for dead reckoning on various robots, including optional contact estimation and gyro filtering to enable low-cost platforms. Extensive experiments on legged, wheeled, and marine robots demonstrate DRIFT's capabilities for long-horizon, perception-free state tracking.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper