Toys, Entertainment robots, Videos Games: Challenges in Design and Control

Transcription

1 Toys, Entertainment robots, Videos Games: Challenges in Design and Control Pranav Bhounsule Department of Mechanical Engineering University of Texas at San Antonio March 23, 2018

2 (1) Toys that walk/run

3 Walking robot: Honda Asimo

4 Wilson Walker

5 Wilson walker (A) (B) (C)

6 Ravert patent (A) (B) (C) UTSA mascot, Rowdy

7 Rowdy Walker

8 Methods & Challenges Leg Design Mass Distribution Integrated Hinge Support Material Commercialization? Time Back leg is connected to body through a hinge Cost Downhill Ramp Front leg is fixed to the body

9 3D printed, linear, ON-OFF, pneumatic actuator (a) (b) (c) (1) Cylinder Head Port A (2) Cylinder Body (6) O-Ring Port B (5) Piston Head (3) Cylinder End Cap (4) Piston Rod

10 Actuator working

11

12 Methods & Challenges (a) Pores (Acetone) (b)strength (Embedding) (c) Piston - Cylinder interface Viton O-rings Waterproof greese

13 Disney s Luxo Jr. Lamp

14 (2) Entertainment Robots

15 Disney animatronics Manually tuned Time consuming

16 Inverse kinematics Bhounsule & Yamane, Humanoids 2015

17 Issues with Kinematics model (a) Head 10 9 (b) Right Hand Left Hand 4 5 B C Z Y X 2 1 A 3 D E Flexible joints -> Rigid body models are invalid Low bandwidth control > poor servo operation High degrees of freedom > Error magnification Wear and Tear > Part/link replacement

18 Iterative Learning Control (ILC ): 1-D example Problem: Move block to the target by applying an instantaneous force Instantaneous force, F Ramp has friction but incorrectly modeled Target

19 Iterative Learning Control (ILC ): 1-D example Model: x = f(f, µ) Imprecise F Target x D Inverse: F = f 1 (x, µ) Imprecise

20 1-D example (trial 1) Control (trial 1): F 1 = f 1 (x, µ) F 1 d 1 e 1 = D d 1 Target x D

21 1-D example (trial 2) Control (trial 2): F 2 = F 1 + e 1 F 2 d2 e 2 = D d 2 Target x D

22 1-D example (trial 2) Converged when e_n is small d n e n = D d n F n Target x D

23 Our approach: Non-linear Inverse Kinematics (IK) update where: i Y i Y ref Y i des ˆF Angle command trial i End-effector position trial i End-effector reference Desired end-effector for IK Estimated Forward Kinematics Model Learning gain

24 Our approach: Non-linear Inverse Kinematics (IK) update Find non-linear IK within joint limits

25 Inverse Kinematics computation Use nonlinear constraint optimization for IK Cost: Bias toward a pose End-effector constraint: Satisfy estimated end-effector position Joint constraint: Satisfy joint limits

26 Inverse kinematics with Iterative Learning Control Bhounsule & Yamane, Humanoids 2015

27 Reference Trial 1 Trial 18 Results for writing task y (m) x (m) Reference Trial 1 Trial 18 Average Square Error, norm (rad 2 ) Iteration Number z (m) Convergence: 18 trials Trial 1: Error ~ 1e-3 Trial 18: Error ~ 1e x (m)

28 Other tasks Bhounsule & Yamane, IJHR 2017

29 (3) Video Games

30 Flappy Bird Game (iphone/android) Control: Tap screen to navigate bird through pipes Scoring: 1 point/pipe passed Objective: Maximize points.

31 Flappy Bird Game History: May 2013: Game released Jan 2014: Most downloaded game on itunes, earning $50,000/day (?) Feb 2014: Game removed from itunes by developer citing its addictive nature

32 Flappy Bird, simple concept but difficult to achieve high score

33 How to beat Flappy Bird downloaded from a YouTuber

34 Past work Machine Learning Reinforcement learning, Q-learning, and Support Vector Machines. Select features, Learn state-action pairs Scores ~

35 Physics Y - vertical height (up -) V - vertical velocity g - gravity (=0.1356) z - control (flap or not flap) constant horizontal velocity

36 #1: Heuristic controller & manual tuning Set-point based control Set-point is tuned. h PipePos1 don t jump jump setpointy c

37 Results: Heuristic controller & manual tuning Average score 56.6/500

38 Results: Heuristic controller & manual tuning

39 #2: Optimization with manually tuned constraints Bounding box constraints Terminal constraints constraint lines NOTE: All dimensions are in pixels bounding box 48 constraints endyhigh h optimized path endvel endylow PipePos PipePos1 prediction horizon

40 #2: Optimization with manually tuned constraints Minimize number of jumps Input: Jump or not (z=0 or 1 resp.) for horizontal distance bet. pipes. Terminal constraints Constraints: constraints 1) Physics (big M method) 2) Bounding box constraint (pipes) 3) Terminal constraints (exit) [3 conditions parameters] optimized path PipePos1 endyhigh h endvel PipePos2 endylow Mixed Integer Programming software Gurobi (intlinprog) prediction horizon

41 Results: Optimization-based control, optimization horizon fixed Perfect score 500/500

42 Results: Optimization-based control, optimization horizon fixed

43 #3: Model Predictive Control Same as #2 but with TWO differences: 1) no terminal constraint constraints 2) prediction horizon (n), control horizon is 1 step. [n is the only free parameter] control horizon optimized path prediction horizon

44 Results: Model Predictive Controller Score Data Best fit line exp(0.004 t^ t ) Computation Average Optimization Time (s) time Data Best fit line exp(0.002 t^ t ) Prediction Horizon Prediction Horizon (x Horizontal distance between pipes) Horizon (x Horizontal distance between pipes) Prediction Horizon

45 Results: Model Predictive Controller (with optimum prediction horizon of 90) Average score 419/500

46 Results: Model Predictive Controller Optimal prediction window is 90 ~1.125 horizontal distance between pipes 55 time steps Key message: Need to plan slightly beyond the next pipe 10 time steps 80 time steps horizontal distance between pipes 90 time steps

47 Results: Model Predictive Controller

48 Discussion Heuristic controller Optimization MPC Score (10 runs, max 500 pipes) Worst case time (sec) ~ Tuning Trial and error tuing Trial and error tuning Can be automated

49 Conclusion Position and speed on exiting the pipe seems to be key factors for good performance Optimization/MPC are too slow for real time implementation MPC best compromise between scores and need for intuitive tuning