Vision-Based Motion Control of Robots Azad Shademan Guest Lecturer CMPUT 412 – Experimental Robotics Computing Science, University of Alberta Edmonton, Alberta, CANADA
Vision-Based Control current desired Left Image Right Image B B A A A Overview of VS, 2. What are the current approaches and their problems), 3. Which problem have we addressed? (what is our motivation for global model estimation) B A. Shademan. CMPUT 412, Vision-based motion control of robots
Vision-Based Control Left Image Right Image B B B A. Shademan. CMPUT 412, Vision-based motion control of robots
Vision-Based Control Feedback from visual sensor (camera) to control a robot Also called “Visual Servoing” Is it any difficult? Images are 2D, the robot workspace is 3D 2D data 3D geometry A. Shademan. CMPUT 412, Vision-based motion control of robots
Where is the camera located? Eye-to-Hand e.g.,hand/eye coordination Eye-in-Hand A. Shademan. CMPUT 412, Vision-based motion control of robots
Visual Servo Control law Position-Based: Robust and real-time pose estimation + robot’s world-space (Cartesian) controller Image-Based: Desired image features seen from camera Control law entirely based on image features A. Shademan. CMPUT 412, Vision-based motion control of robots
Position-Based Desired pose Estimated pose A. Shademan. CMPUT 412, Vision-based motion control of robots
Image-Based Desired Image feature Extracted image feature A. Shademan. CMPUT 412, Vision-based motion control of robots
Visual-motor Equation x1 x2 x3 x4 q=[q1 … q6] Visual-Motor Equation This Jacobian is important for motion control. A. Shademan. CMPUT 412, Vision-based motion control of robots
Visual-motor Jacobian Image space velocity Joint space velocity A B A. Shademan. CMPUT 412, Vision-based motion control of robots
Image-Based Control Law Measure the error in image space Calculate/Estimate the inverse Jacobian Update new joint values A. Shademan. CMPUT 412, Vision-based motion control of robots
Image-Based Control Law Desired Image feature Extracted image feature A. Shademan. CMPUT 412, Vision-based motion control of robots
Jacobian calculation Analytic form available if model is known. Known model Calibrated Must be estimated if model is not known Unknown model Uncalibrated A. Shademan. CMPUT 412, Vision-based motion control of robots
Image Jacobian (calibrated) Analytic form depends on depth estimates. Camera Velocity Camera/Robot transform required. No flexibility. A. Shademan. CMPUT 412, Vision-based motion control of robots
Image Jacobian (uncalibrated) A popular local estimator: Recursive secant method (Broyden update): A. Shademan. CMPUT 412, Vision-based motion control of robots
Calibrated vs. Uncalibrated Model derived analytically Global asymptotic stability Optimal planning is possible A lot of prior knowledge on the model Relaxed model assumptions Traditionally: Local methods No global planning Difficult to show asymptotic stability condition is ensured The problem of traditional methods is the locality. Global Model Estimation (Research result) Optimal trajectory planning Global stability guarantee Kinematics A. Shademan. CMPUT 412, Vision-based motion control of robots
Synopsis of Global Visual Servoing Model Estimation (Uncalibrated) Visual-Motor Kinematics Model Global Model Extending Linear Estimation (Visual-Motor Jacobian) to Nonlinear Estimation Our contributions: K-NN Regression-Based Estimation Locally Least Squares Estimation A. Shademan. CMPUT 412, Vision-based motion control of robots
Local vs. Global 1st Rank Broyden update: Jägersand et al. ’97 Key idea: using only the previous estimation to estimate the Jacobian RLS with forgetting factor Hosoda and Asada ’94 1st Rank Broyden update: Jägersand et al. ’97 Exploratory motion: Sutanto et al. ‘98 Quasi-Newton Jacobian estimation of moving object: Piepmeier et al. ‘04 Key idea: using all of the interaction history to estimate the Jacobian Globally-Stable controller design Optimal path planning Local methods don’t! RLS with forgetting factor, 1st rank broyden, exploratory motion, quasi-Newton Jacobian estimation of a moving object A. Shademan. CMPUT 412, Vision-based motion control of robots
K-NN Regression-based Method ? q1 q2 x1 q2 3 NN q1 A. Shademan. CMPUT 412, Vision-based motion control of robots
Locally Least Squares Method ? q1 q2 x1 (X,q) KNN(q) A. Shademan. CMPUT 412, Vision-based motion control of robots
Experimental Setup Puma 560 Eye-to-hand configuration Stereo vision Features: projection of the end-effector’s position on image planes (4-dim) 3 DOF for control A. Shademan. CMPUT 412, Vision-based motion control of robots
Measuring the Estimation Error Hidden? A. Shademan. CMPUT 412, Vision-based motion control of robots
Global Estimation Error Local estimation and KNN have the same order, but the variance of KNN method is much less. A. Shademan. CMPUT 412, Vision-based motion control of robots
Noise on Estimation Quality KNN LLS With increasing noise level, the error decreases A. Shademan. CMPUT 412, Vision-based motion control of robots
Effect of Number of Neighbors A. Shademan. CMPUT 412, Vision-based motion control of robots
Conclusions Presented two global methods to learn the visual-motor function LLS (global) works better than the KNN (global) and local updates. KNN suffers from the bias in local estimations Noise helps system identification A. Shademan. CMPUT 412, Vision-based motion control of robots
Eye-in-Hand Simulator A. Shademan. CMPUT 412, Vision-based motion control of robots
Eye-in-Hand Simulator A. Shademan. CMPUT 412, Vision-based motion control of robots
Eye-in-Hand Simulator A. Shademan. CMPUT 412, Vision-based motion control of robots
Eye-in-Hand Simulator A. Shademan. CMPUT 412, Vision-based motion control of robots
Mean-Squared-Error A. Shademan. CMPUT 412, Vision-based motion control of robots
Task Errors A. Shademan. CMPUT 412, Vision-based motion control of robots
Questions? A. Shademan. CMPUT 412, Vision-based motion control of robots
Position-Based Robust and real-time relative pose estimation Extended Kalman Filter to solve the nonlinear relative pose equations. Cons: EKF is not the optimal estimator. Performance and the convergence of pose estimates are highly sensitive to EKF parameters. A. Shademan. CMPUT 412, Vision-based motion control of robots
2D-3D nonlinear point correspondences Overview of PBVS 2D-3D nonlinear point correspondences What kind of nonlinearity? IEKF T. Lefebvre et al. “Kalman Filters for Nonlinear Systems: A Comparison of Performance,” Intl. J. of Control, vol. 77, no. 7, pp. 639-653, May 2004. A. Shademan. CMPUT 412, Vision-based motion control of robots
Measurement equation is nonlinear and must be linearized. EKF Pose Estimation yaw pitch roll State variable Process noise Measurement noise Measurement equation is nonlinear and must be linearized. A. Shademan. CMPUT 412, Vision-based motion control of robots
Visual-Servoing Based on the Estimated Global Model A. Shademan. CMPUT 412, Vision-based motion control of robots
Control Based on Local Models See Spong etc book. See if this should be left out as “hidden” A. Shademan. CMPUT 412, Vision-based motion control of robots
Estimation for Local Methods We need to estimate the Jacobian which is equal to minimizing the following problem. In fact we want to fit a plan to the local neighbourhood of the current point q. In discrete form, … This is not how we estimate the Jacobian in practice, rather we estimate the Jacobian locally, that is, at each point we fit a plane to the nonlinear model. For example, by Broyden’s first order method, we estimate the Jacobian as, or using the RLS with forgetting factor (see the paper) In practice: Broyden 1st-rank estimation, RLS with forgetting factor, etc. A. Shademan. CMPUT 412, Vision-based motion control of robots
A. Shademan. CMPUT 412, Vision-based motion control of robots