Runbot

Team members: Marie Bro Duun Georgious Evangelos Emre Ozbilge Antonio Gomez Zamorano Matej Hoffmann

Supervisor: Tao Geng

Goal

Make the Runbot robot learn to adjust step length Parameters:

Maximum voltage to hip motors Extreme angle of hip joint

AEP – anterior extreme position

Emergency goal: make the robot walk without a touch sensor

Relationship between parameters

Step length = f(max voltage, hip angle) ? A nontrivial nonlinear relationship Stability issue

4 5 6 7 8 9 10 11

0

2

4

6

8

10

12

14

16

18

20

Hip min angle 74

Column A

Hip max voltage

Ste

p le

ngth

Optimization algorithms

'Heuristic' Evolutionary algorithms Simulated annealing

Gradient ascent methods Methods with memory – e.g. Q-learning

Achievements

Real robot: Going to target step length from some initial

conditions

Simulation Optimization algorithm testbed – Simulink Several gradient based optimization methods tailored

to the problem

Algorithm test bed

Pseudocode:

1) Short/Long Term Error Estimation

2) Relate Delta Constant to Estimated Error

3) Parameter selection by randomization

4) Parameter learning for Short/Long Term Gradient Policy Approach

Open questions

Fitness landscape Can gradient be obtained reliably? Are there too many local minima? Fitness vs. stability Other control parameters? Step length vs. speed

Next steps

Obtain a systematic rough picture of the fitness landscape from the real robot to assess feasibility of different optimization methods (e.g. gradient vs. non-gradient, methods with memory...)

Create a similar landscape in testbed and compare algorithms

Run experiments on real robot

Thank you for your attention!