State Estimation for Autonomous Vehicles Stergios I. Roumeliotis Computer Science & Engineering Department University of Minnesota stergios@cs.umn.edu www.cs.umn.edu/~stergios
Outline Sensing & Estimation Estimator Requirements Indirect Kalman filter formulation SC-KF Preliminary results Ongoing & related work Challenges & unresolved issues Extensions & future work
Sensors Proprioceptive: Directly measure the motion of the vehicle IMU (accels. & gyros) Doppler radar Noise Integration Exteroceptive: Measure “identities” of the environment, or, “relation” of vehicle with the environment. Used to infer absolute and/or relative position and orientation (pose displacement) Compass, Sun Sensor GPS Cameras (single, stereo, omni, FLIR) Laser scanner, MW radar, Sonar Wheel Encoders … Uncertainty & Noise
State Estimation Propagation Update Techniques Bayesian estimation Kalman filter Particle filter Unscented filter … Goal: Estimate & Control State of vehicle (position, orientation, velocity, direction of motion, …) State of environment (detect obstacles, position of objects of interest, area identities, mapping, …) Propagation Update
Estimator Requirements Portable (independent of vehicle) Adaptable (number & type of sensors) Modular (robustness against single point sensor failures) Time flexible (able to process synchronous & asynchronous sensor measurements) Expandable to multi-robot systems
Sensor &Vehicle Independence Adaptability & Portability: Estimator considers any vehicle as a static network of sensors at known configuration
Indirect Kalman filter – Sensor Modeling State Propagation: Integrate sensor measurements from these of the sensors that measure highest order derivatives of motion When IMU part of sensor payload, integrates Advantages, compared to Dynamic Modeling Difficult to derive precise vehicle/environment dynamics Vehicle modifications require new derivation CPU cost (large state vector to capture dynamics) Statistical Modeling (commonly used for target tracking) Motion statistics unknown/uncertain State Update: Asynchronously when new measurements available.
Indirect Kalman filter - Formulation Quantities of interest State vector Estimated State vector Error state vector Propagation Continuous time error state propagation Covariance propagation
Indirect Kalman filter – Update (1 time instant) Measurement is a function of state vector at a certain time instant Position (GPS, UHF link, DSN) Orientation (Compass, sun sensor) Linear Velocity (Doppler radar) Observer (Estimator) Controller H MATRICES FOR CATEGORY 1
Indirect Kalman filter – Update (2 time instants) Measurement is a function of state vector at more than one time instant Estimated Rotational & Translational Displacement (Relative State Measurement) Visual odometry (mono, stereo) Laser scan matching Kinematics-based vehicle odometry
Example: Weighted Laser Scan Matching Relative position and orientation measurement inferred by correlating sensor measurements recorded at 2 separate locations.
State Estimation & Relative Pose Measurements Propagation Sensor Model Proprioceptive Measurements (“continuously”) Update Sensor Models Exteroceptive Measurements (intermittently)
Previous Approaches I 1. Approximate as higher order derivatives
Previous Approaches II 2. Approx. as absolute state pseudo-measurement [Hoffman, Baumgartner, Huntsberger, Shenker ’99] 3. Estimate relative states instead (2 estimators) FILTER
Stochastic Cloning –Kalman Filter (SC-KF) Relative State Measurement Relative State Measurement Error Augmented State Vector
SC-KF Propagation Equations State Propagation Augmented Error State & Covariance Augmented Error State & Covariance Propagation
SC-KF Update Equations Residual Covariance
Estimation Block Diagram (Helicopter) Estimators Sensors Camera IMU 3 accelerometers & 3 gyroscopes Laser Altimeter SC- KF Inertial Sensor Integrator Kalman filter Visual Feature Tracking pixel images distance to features
Preliminary Results – Experimental Setup Average absolute errors in p = [x y z]: IMU alone [53.5 464.7 126.1] mm (not shown on Fig. due to errors magnitude) VISION alone [17.4 41.4 29.9] mm KF: IMU & VISION [4.5 4.7 4.2] mm simulated planetary surface helicopter E-Box
Preliminary Results - W/out sensor fusion
Preliminary Results - Altitude & Bias Estimates
Experiments w/ Mobile Robot I Wheel Odometry and Weighted Laser Scan Matching
Experiments w/ Mobile Robot I Total Distance: 22.25 mm Average Distance Errors Odometry: 258 mm WLSM: 95 mm SC-KF: 77 mm
State Covariance - Simulation
Ongoing & Related Work Treat time delays of vision algorithms (e.g. visual odometry) SC2-KF (3 copies of the state) Detect kinematics-based odometry errors Slippage Estimation Smoother – Trajectory Reconstruction Attitude estimation between consecutive stops of the rover
Ongoing Work - *Unresolved Issues* Sensor Alignment Determine 3D transformation between pairs of sensors Must be accurate to correlate sensor measurements w/out errors Tedious & time consuming process when done manually Active Sensor Alignment Determine motions that excite all d.o.f. and allow sensor network on the vehicle body to self-configure
Extensions & Future Work Extension to Simultaneous Localization And Mapping (SLAM) Incorporate, update, and enhance previous maps of area Satellite imagery, EDL Challenges: Computational complexity O(N2) Proposed solution: FWPT compression of covariance matrix P Fault detection and identification Structural damages Sensor failures Distributed state estimation Reconfigurable, mobile networks of robots & sensors
Acknowledgements DARPA, Tactical Mobile Robot Program (JPL) Cog: Robert Hogg, PI: Larry Matthies NASA Ames, IS program (JPL) “Safe & Precise Landing,” Cog: Jim Montgomery, PI: Larry Matthies NASA Mars Technology Program, (JPL) “Navigation on Slopes,” Cog: Dan Helmick, PI: Larry Matthies “CLARAty”, PI: Issa Nesnas University of Minnesota (UMN), GIA program PI: Stergios Roumeliotis NSF, ITR program (UMN) PI: Nikos Papanikolopoulos NSF, Ind./Univ. Cooperative Research Center (UMN) PI: Richard Voyles