ADVANCING ACTIVE VISION BY IMPROVED DESIGN AND CONTROL

1. Robotic Systems Laboratory ADVANCING ACTIVE VISIONBY IMPROVEDDESIGN AND CONTROL The Australian National University Research School of Information Sciences and Engineering Orson Sutherland, Harley Truong, Sébastien Rougeaux & Alexander Zelinsky

2. Introduction • RSL has developed an extensive library of Vision Processing software for : • Human/Machine interfaces • Human/Robot interaction • Navigation

3. Introduction • The intention is to mount these products into a powerful and robust sensor. • For this we have chosen Active Vision.

4. Exisiting Devices The Agile Eye ESCHeR

5. Introduction • This presentation will concentrate on: • the mechanical design • the Hardware • Performance Specifications and Testing • and the Control of a novel active Vision system named CeDAR for CableDrive Active Vision

6. CeDAR

7. Mechanical Design • We wanted: • a high performance • high precision rig • capable of moving a variety of payloads • For this we required: • a high precision transmission • and an optimised mechanical structure

8. Mechanical Design Issues • Speed & acceleration depend on: • motor size • load inertia • transmission ratio • Accuracy depends on: • encoder resolution • transmission ratio and stiffness • joint stiffness

9. Kinematics • CeDAR is arranged in the more popular Helmholtz configuration. • An important property of the design is that the axes intersect at the optical center of each camera Helmholtz Fick

10. Transmission System • Uses a cable drive transmission inspired by the Whole Arm Manipulator (Townsend, MIT) • Same as gear transmission except force is transmitted by tension in cables rather than contact between teeth

11. Transmission System • Advantages: • No Backlash • No Slippage • No lubrication • High efficiency • No speed limits • Torque limited only by strength of cables Bevel

12. Transmission System • Disadvantages: • Limited range. • Miniaturisation is limited by the minimum bend radius of the cables. HOWEVER • In our system, range is only 90°. • Pulleys are integrated into structure.

13. Mechanical Architecture • Parallel architecture allows: • Motors to be fixed at base so they do not contribute to mass in motion • But means axes are coupled

14. Optimisation The weight/rigidity trade-off was optimised so that maximum angular deflection was 0.01°.

15. Hardware Overview • Fully assembled CeDAR weighs 3.5kg. • It has a moving mass of 1.7kg including 2x350g digital cameras. • Maxon DC motors controlled by Motion Engineering, motion card. • Pentium III performs trajectory planning and vision processing

16. Performance Specs • The max range, payload and baseline were based on the use of motorised-zoom cameras (350g). • Performance needed to be at least comparable to the human eye: • 5x90° saccades per second

17. Performance Testing • We achieved: • Tilt: 600°.s-1 @ 18000°.s-2 with 0.01° and a saccade rate of 5Hz. • Vergence: 800°.s-1 @ 20000°.s-2 with 0.01° and a saccade rate of 6Hz. • This means: • Better performance than the human eye. • Better performance than ESCHeR. • Marginally inferior to the Agile Eye.

18. Control • Extends work by Murray et al. (1992) on Trapezoidal Profile Motion (TPM). • In theory TPM can achieve optimal saccade and smooth pursuit with one compact algorithm.

19. Control Position profiles for saccade and smooth pursuit. Velocity profiles for saccade and Smooth pursuit.

20. Control Two successive real-time saccades Smooth pursuit of a 0.5Hz sinusoid, sampled at 4Hz.

21. Future Work • Applications: • Tracking with ZDF and Optic Flows • Face/Feature tracking algorithms • Hardware Improvements: • A pan neck • Future Prototypes: • Use of plastics and other polymers • Fick configuration

22. Conclusions • CeDAR is a fast, accurate 3DOF stereo head. • CeDAR can carry relatively heavy payloads. • CeDAR’s controller is compact and simple and can implement both optimal saccade and smooth pursuit from the same algorithm.