Discriminative Training of Kalman Filters P. Abbeel, A. Coates, M

Slides:



Advertisements
Similar presentations
Mobile Robot Localization and Mapping using the Kalman Filter
Advertisements

EKF, UKF TexPoint fonts used in EMF.
Adam Coates, Pieter Abbeel, and Andrew Y. Ng Stanford University ICML 2008 Learning for Control from Multiple Demonstrations TexPoint fonts used in EMF.
Learning Parameterized Maneuvers for Autonomous Helicopter Flight Jie Tang, Arjun Singh, Nimbus Goehausen, Pieter Abbeel UC Berkeley.
IR Lab, 16th Oct 2007 Zeyn Saigol
Introduction to Mobile Robotics Bayes Filter Implementations Gaussian filters.
Probabilistic Robotics: Kalman Filters
Attitude Determination - Using GPS. 20/ (MJ)Danish GPS Center2 Table of Contents Definition of Attitude Attitude and GPS Attitude Representations.
Using Inaccurate Models in Reinforcement Learning Pieter Abbeel, Morgan Quigley and Andrew Y. Ng Stanford University.
Stanford CS223B Computer Vision, Winter 2007 Lecture 12 Tracking Motion Professors Sebastian Thrun and Jana Košecká CAs: Vaibhav Vaish and David Stavens.
Introduction to Kalman Filter and SLAM Ting-Wei Hsu 08/10/30.
Mobile Intelligent Systems 2004 Course Responsibility: Ola Bengtsson.
Probabilistic video stabilization using Kalman filtering and mosaicking.
CS 547: Sensing and Planning in Robotics Gaurav S. Sukhatme Computer Science Robotic Embedded Systems Laboratory University of Southern California
7. Experiments 6. Theoretical Guarantees Let the local policy improvement algorithm be policy gradient. Notes: These assumptions are insufficient to give.
Probabilistic Robotics
Stanford CS223B Computer Vision, Winter 2007 Lecture 12 Tracking Motion Professors Sebastian Thrun and Jana Košecká CAs: Vaibhav Vaish and David Stavens.
Maximum Likelihood (ML), Expectation Maximization (EM)
Estimation and the Kalman Filter David Johnson. The Mean of a Discrete Distribution “I have more legs than average”
Course AE4-T40 Lecture 5: Control Apllication
Probabilistic Robotics Bayes Filter Implementations Gaussian filters.
Novel approach to nonlinear/non- Gaussian Bayesian state estimation N.J Gordon, D.J. Salmond and A.F.M. Smith Presenter: Tri Tran
Our acceleration prediction model Predict accelerations: f : learned from data. Obtain velocity, angular rates, position and orientation from numerical.
Kalman Filtering Pieter Abbeel UC Berkeley EECS Many slides adapted from Thrun, Burgard and Fox, Probabilistic Robotics TexPoint fonts used in EMF. Read.
Overview and Mathematics Bjoern Griesbach
Particle Filtering. Sensors and Uncertainty Real world sensors are noisy and suffer from missing data (e.g., occlusions, GPS blackouts) Use sensor models.
ROBOT MAPPING AND EKF SLAM
Kalman filter and SLAM problem
Lecture 11: Kalman Filters CS 344R: Robotics Benjamin Kuipers.
Kalman filtering techniques for parameter estimation Jared Barber Department of Mathematics, University of Pittsburgh Work with Ivan Yotov and Mark Tronzo.
Colorado Center for Astrodynamics Research The University of Colorado 1 STATISTICAL ORBIT DETERMINATION Satellite Tracking Example of SNC and DMC ASEN.
Computer vision: models, learning and inference Chapter 19 Temporal models.
3D SLAM for Omni-directional Camera
Kalman Filter (Thu) Joon Shik Kim Computational Models of Intelligence.
Jamal Saboune - CRV10 Tutorial Day 1 Bayesian state estimation and application to tracking Jamal Saboune VIVA Lab - SITE - University.
Probabilistic Robotics Bayes Filter Implementations Gaussian filters.
Robust localization algorithms for an autonomous campus tour guide Richard Thrapp Christian Westbrook Devika Subramanian Rice University Presented at ICRA.
Young Ki Baik, Computer Vision Lab.
Human-Computer Interaction Kalman Filter Hanyang University Jong-Il Park.
ECE 8443 – Pattern Recognition ECE 8423 – Adaptive Signal Processing Objectives: ML and Simple Regression Bias of the ML Estimate Variance of the ML Estimate.
Processing Sequential Sensor Data The “John Krumm perspective” Thomas Plötz November 29 th, 2011.
A Passive Approach to Sensor Network Localization Rahul Biswas and Sebastian Thrun International Conference on Intelligent Robots and Systems 2004 Presented.
Real-Time Simultaneous Localization and Mapping with a Single Camera (Mono SLAM) Young Ki Baik Computer Vision Lab. Seoul National University.
State Estimation and Kalman Filtering
NCAF Manchester July 2000 Graham Hesketh Information Engineering Group Rolls-Royce Strategic Research Centre.
An Introduction To The Kalman Filter By, Santhosh Kumar.
Particle Filtering. Sensors and Uncertainty Real world sensors are noisy and suffer from missing data (e.g., occlusions, GPS blackouts) Use dynamics models.
By: Aaron Dyreson Supervising Professor: Dr. Ioannis Schizas
Extended Kalman Filter
Nonlinear State Estimation
Cameron Rowe.  Introduction  Purpose  Implementation  Simple Example Problem  Extended Kalman Filters  Conclusion  Real World Examples.
Particle Filtering. Sensors and Uncertainty Real world sensors are noisy and suffer from missing data (e.g., occlusions, GPS blackouts) Use sensor models.
Tracking Mobile Nodes Using RF Doppler Shifts
The Unscented Particle Filter 2000/09/29 이 시은. Introduction Filtering –estimate the states(parameters or hidden variable) as a set of observations becomes.
A Low-Cost and Fail-Safe Inertial Navigation System for Airplanes Robotics 전자공학과 깡돌가
Colorado Center for Astrodynamics Research The University of Colorado 1 STATISTICAL ORBIT DETERMINATION Kalman Filter with Process Noise Gauss- Markov.
The Unscented Kalman Filter for Nonlinear Estimation Young Ki Baik.
Robust Localization Kalman Filter & LADAR Scans
Camera calibration from multiple view of a 2D object, using a global non linear minimization method Computer Engineering YOO GWI HYEON.
Beard & McLain, “Small Unmanned Aircraft,” Princeton University Press, 2012, Chapter 8, Slide 1 Chapter 8 State Estimation.
Probabilistic Robotics Bayes Filter Implementations Gaussian filters.
Using Sensor Data Effectively
Variational filtering in generated coordinates of motion
ASEN 5070: Statistical Orbit Determination I Fall 2014
Velocity Estimation from noisy Measurements
PSG College of Technology
Course: Autonomous Machine Learning
Motion Models (cont) 2/16/2019.
Bayes and Kalman Filter
Kalman Filtering COS 323.
Presentation transcript:

Discriminative Training of Kalman Filters P. Abbeel, A. Coates, M Discriminative Training of Kalman Filters P. Abbeel, A. Coates, M. Montemerlo, A. Y. Ng, S. Thrun Kalman filters estimate the state of a dynamical system from inputs and measurements. The Kalman filter’s parameters (e.g., state and observation variances) significantly affect its accuracy, and are difficult to choose. Current practice: hand-engineer the parameters to get the best possible result.. We propose to collect ground-truth data and then learn the Kalman filter parameters automatically from the data using discriminative training.

Discriminative Training of Kalman Filters P. Abbeel, A. Coates, M Discriminative Training of Kalman Filters P. Abbeel, A. Coates, M. Montemerlo, A. Y. Ng, S. Thrun Ground truth Hand-engineered Learned

Discriminative Training of Kalman Filters Pieter Abbeel, Adam Coates, Mike Montermerlo, Andrew Y. Ng, Sebastian Thrun Stanford University

Motivation Extended Kalman filters (EKFs) estimate the state of a dynamical system from inputs and measurements. For fixed inputs and measurements, the estimated state sequence depends on: Next-state function. Measurement function. Noise model. Don’t focus too soon on noise terms. State problem at more general level. This paper: Focus on covariance.

Motivation (2) Noise terms typically result from a number of different effects: Mis-modeled system dynamics and measurement dynamics. The existence of hidden state in the environment not modeled by the EKF. The discretization of time. The algorithmic approximations of the EKF itself, such as the Taylor approximation commonly used for linearization. In Kalman filters noise is assumed independent over time, in practice noise is highly correlated.

Motivation (3) Common practice: careful hand-engineering of the Kalman filter’s parameters to optimize its performance, which can be very time consuming. Proposed solution: automatic learning of the Kalman filter’s parameters from data where (part of) the state-sequence is observed (or very accurately measured). In this work: we focus on learning the state and measurement variances, although the principles are more generally applicable.

Example Problem: estimate the variance of a GPS unit used to estimate the fixed position x of a robot. Standard model: xmeasured = xtrue +   » N(0,2) (Gaussian with mean 0, variance 2) Assuming the noise is independent over time, we have after n measurements variance = 2/n. However if the noise is perfectly correlated, the true variance is 2 >> 2/n. Practical implication: wrongly assuming independence leads to overconfidence in the GPS sensor. This matters not only for the variance, but also for the state estimate when information from multiple sensors is combined.

Extended Kalman filter State transition equation: xt = f(xt-1,ut) +   » N(0, R) (Gaussian with mean zero, covariance R) Measurement equation: zt = g(xt) +   » N(0,Q) (Gaussian with mean zero, covariance Q) The extended Kalman filter linearizes the non-linear function f, g through their Taylor approximation: f(xt-1,ut) ¼ f(t-1,ut) + Ft(xt-1-t-1) g(xt) ¼ g(t) + Gt(xt - t) Here Ft and Gt are Jacobian matrices of f and g respectively, taken at the filter estimate .

Extended Kalman Filter (2) For linear systems, the (standard) Kalman filter produces exact updates of expected state and covariance. The extended Kalman filter applies the same updates to the linearized system. Prediction update step: Place holder Measurement update step:

Discriminative training Let y be a subset of the state variables x for which we obtain ground-truth data. E.g., the positions obtained with a very high-end GPS receiver that is not part of the robot system when deployed. We use h(.) to denote the projection from x to y: y = h(x). Discriminative training: Given (u1:T, z1:T, y1:T). Find the filter parameters that predict y1:T most accurately. [Note: it is actually sufficient that y is a highly accurate estimate of x. See the paper for details.]

Three discriminative training criteria Minimizing the residual prediction error: Maximizing the prediction likelihood: Maximizing the measurement likelihood: (The last criterion does not require access to y1:T.)

Evaluating the training criteria The extended Kalman filter computes p(xt|z1:t,u1:t) = N(t, t) for all times t 2 {1 … T}. Residual prediction error and prediction likelihood can be evaluated directly from the filter’s output. Measurement likelihood:

Robot testbed

The robot’s state, inputs and measurements x,y : position coordinates. : heading. v : velocity. b : gyro bias. Control inputs: r : rotational velocity. a : forward acceleration. Measurements: Optical wheel encoders measure v. A (cheap) GPS unit measures x,y (1Hz, 3m accuracy). A magnetic compass measures .

System Equations

Experimental setup We collected two data sets (100 seconds each) by driving the vehicle around on a grass field. One data set is used to discriminatively learn the parameters, the other data set is used to evaluate the performance of the different algorithms. A Novatel RT2 differential GPS unit (10Hz, 2cm accuracy) was used to obtain ground truth position data. Note the hardware on which our algorithms are evaluated do not have the more accurate GPS.

Experimental results (1) Evaluation metrics: RMS error (on position): Prediction log-loss (on position): Method Residual Error Prediction Likelihood Measurement Likelihood CMU hand- tuned RMS error 0.26 0.29 0.50 log-loss -0.23 0.048 40 0.75

Experimental results (2) Zoomed in on this area on next figure. Ground truth CMU hand-engineered Learned minimizing residual prediction error

Experimental results (3) Ground truth CMU hand-engineered Learned minimizing residual prediction error

Experimental Results (4) Zoomed in on this area on next figure. x (cheap) GPS Ground truth CMU hand-engineered Learned minimizing residual prediction error Learned maximizing prediction likelihood measurement likelihood

Experimental Results (5) x (cheap) GPS Ground truth CMU hand-engineered Learned minimizing residual prediction error Learned maximizing prediction likelihood Learned maximizing measurement likelihood

Related Work Conditional Random Fields (J. Lafferty, A. McCallum, F. Pereira, 2001). The optimization criterion (translated to our setting) is: Issues for our setting: The criterion does not use filter estimates. (It uses smoother estimates instead.) The criterion assumes the state is fully observed.

Discussion and conclusion In practice Kalman filters often require time-consuming hand-engineering of the noise parameters to get optimal performance. We presented a family of algorithms that use a discriminative criterion to learn the noise parameters automatically from data. Advantages: Eliminates hand-engineering step. More accurate state estimation than even a carefully hand-engineered filter.

Discriminative

Training of

Kalman Filters

Pieter Abbeel, Adam Coates, Mike Montemerlo, Andrew Y Pieter Abbeel, Adam Coates, Mike Montemerlo, Andrew Y. Ng and Sebastian Thrun Stanford University

Pieter Abbeel, Adam Coates, Mike Montemerlo, Andrew Y Pieter Abbeel, Adam Coates, Mike Montemerlo, Andrew Y. Ng and Sebastian Thrun Stanford University

Robot testbed