State Estimation

Estimates a robot’s current position, orientation, and velocity from noisy sensor readings.

State estimation infers quantities that can’t be measured directly — such as body pose, velocity, and sensor bias — from sensor measurements and a motion model. Every sensor has noise and drift, so readings must be fused according to how much each can be trusted, typically using the Kalman filter family, particle filters, or factor-graph optimization. Take a legged robot as an example: the IMU measures angular velocity and acceleration, joint encoders combined with leg kinematics give each foot’s position relative to the body, and assuming the stance foot doesn’t slip lets the system estimate body velocity. A 2018 University of Michigan paper validated this approach with a contact-aided invariant EKF on the Cassie biped. The body linear velocity that reinforcement-learning locomotion policies need comes from a state estimator; some work instead trains a neural network to learn that estimate directly.

ExampleWhile a quadruped robot walks blind, IMU readings, joint angles, and foot contact states are fed into an extended Kalman filter every control cycle, which outputs body orientation and linear velocity for the locomotion policy to use.

Also called

Robot State Estimation

Related

Kalman Filter · Invariant Extended Kalman Filter · Leg Odometry · Inertial Measurement Unit · Proprioception · Learned State Estimator

Sources

Timothy Barfoot 主页：State Estimation for Robotics（第二版 2024，含免费 PDF 与中译本信息） (Chinese)
Contact-Aided Invariant Extended Kalman Filtering for Legged Robot State Estimation (RSS 2018)