1. Robot Intelligence Technology Lab. Evolution of simple navigation Chapter 4 of Evolutionary Robotics Jan. 12, 2007 YongDuk Kim

2. 2 Robot Intelligence Technology Lab. Contents Introduction Straight motion with obstacle avoidance Visually guided navigation Re-adaptation Cross platform adaptation From simulation to reality Conclusions

3. 3 Robot Intelligence Technology Lab. 1. Introduction In this chapter, evolution of simple behaviors will be described. Navigation ability The development of a suitable mapping from sensory information to motor actions. The closed feedback loop (between sensory information and motor actions) makes it rather difficult to design a stable control system for realistic situations. One solution Listing all possible sensory situations and associate them to a set of predefined motor actions. The solution is not always viable because of unknown and unpredictable environments. Artificial evolution Smart controller by exploiting interactions between the robot and the environment.

4. 4 Robot Intelligence Technology Lab. 2. Straight motion with obstacle avoidance Braitenberg’s vehicle [Braitenberg 1984] The robot morphology is symmetrical and it has two wheels. It is conceptual robot whose wheels are directly linked to the sensors through weighted connections. Hand designed solution It should be noticed that even this simple design requires careful analysis of sensor and motor profiles, and important decisions for what concerns the direction of motion, its straightness, and its velocity. Different robots and different environments require different set of carefully chosen values.

5. 5 Robot Intelligence Technology Lab. 2. Straight motion with obstacle avoidance Evolutionary approach [Floreano and Mondada 1994] It could find a solution for straight navigation and obstacle avoidance without assuming all the prior knowledge about sensors, motors, and environment. The goal was to evolve a control system capable of maximizing forward motion while accurately avoiding all obstacles on its way. The fitness function Where V is the sum of rotation speed of the two wheels, Δv is the absolute value of the algebraic difference between the signed speed values of the wheels, i is the normalized activation value of the infrared sensor with the highest activity.

6. 6 Robot Intelligence Technology Lab. 2. Straight motion with obstacle avoidance These three components encourage – respectively – motion, straight displacement, and obstacle avoidance, but do not say in what direction the robot should move. The control system A neural network  One layer of synaptic weights from the eight infrared sensors to two motors units. The results

7. 7 Robot Intelligence Technology Lab. 2. Straight motion with obstacle avoidance For each generation, it took approximately 40 minutes. Although the fitness indicators keep growing for 100 generations, around the 50 th generation the best individuals already exhibited a smooth navigation around the maze. A fitness value of 1.0 could have been achieve only by a robot moving straight at maximum speed in an open space. In the experiments, 0.3 was the maximum value attained by the evolutionary controller even when continued for further 100 generations.

8. 8 Robot Intelligence Technology Lab. 2. Straight motion with obstacle avoidance Values of the fitness components

9. 9 Robot Intelligence Technology Lab. 3. Visually guided navigation It is very likely that by the end of the this decade almost all autonomous robots will employ vision as a primary sensory system. Mainstream approach to vision processing [Marr 1982] Based on preprocessing, segmentation, and pattern recognition Is not viable for systems that must respond very quickly in partially unpredictable environments. A drastic new approach Takes into account the ecological aspects of visual based behavior and its integration with the motor system. There were only few efforts into this direction. [Horswill 1993; Marjanovic et al. 1996]

10. 10 Robot Intelligence Technology Lab. 3. Visually guided navigation Evolutionary robotics provides an ideal framework. It allows the development of visual processing along with motor processing in closed feedback loop. It relies less on externally imposed assumption. It allows simultaneous exploration of both controllers and sensor morphologies. The ecological vision is going to be a very fertile area for evolutionary robotics over the next years.

11. 11 Robot Intelligence Technology Lab. 3. Visually guided navigation Gantry robot [Harvey 192a, 1992b, 1993]

12. 12 Robot Intelligence Technology Lab. 3. Visually guided navigation The visual input is considerably reduce by sampling only a small part of the image according to genetically specified instructions. The neural networks have a fixed number of input and outputs Artificial evolution was carried out in three stages of increasing behavioral complexity

13. 13 Robot Intelligence Technology Lab. 3. Visually guided navigation

14. 14 Robot Intelligence Technology Lab. 4. Re-adaptation The price to pay for the automatics process of adaptation described above is the amount of time required by evolution carried out entirely on physical robots. The question the is to what extent and at what speed an evolutionary system can generalize and/or re-adapt to modified environmental conditions without retraining it from scratch. Cross platform adaptation [Floreano and Mondada 1998] In some cases, it might be desirable to continue evolution on the new robot incrementally. From the point of view of the neurocontroller, changing the sensory motor characteristics of the robot is just another way of modifying the environment.

15. 15 Robot Intelligence Technology Lab. 4. Re-adaptation Incremental evolution still requires quite a lot of research in order to accommodate more complex sensory motor systems, acquisition of new skills, modification of old ones.

16. 16 Robot Intelligence Technology Lab. 4. Re-adaptation From simulation to reality Simulations can provide a valuable aid to evolutionary robotics as long as they are coupled with test s on physical robots. Transferring an evolved controller from the simulated to the real robot is very likely to generate discrepancies in the behavior and performance of the robot caused by different properties of the sensory motor interactions between the robot and its environment. Experiments [Miglino et al., 1995]  No-noise condition: not including any noise  Noise condition: adding uniform white noise to the simulated sensors.  Conservative-noise condition: the sensory values were read as if the robot had been displaced by a small random quantity along the x and y coordinates.

17. 17 Robot Intelligence Technology Lab. 4. Re-adaptation Environments

18. 18 Robot Intelligence Technology Lab. 5. Conclusions Some examples of artificial evolution applied to simple navigation tasks is presented. The results indicate that artificial evolution can be fruitfully applied even to tasks where a preprogrammed strategy already exists or to tasks that are apparently simple. Even slight modifications to the environment of an evolved individual are likely to cause a drop in performance. Performance can be rapidly recovered.