|
3.2 Transfer to the Real Robot
A way of verifying the evolutionary robotics results obtained in simulation is to
test the evolved neural controllers on a real robot. For this purpose the T-Maze
shown in figure 7 was built. The best individual from each of the 10 replications
of experiment 1 task was tested. Initially however, the results of the tests
were rather poor. None of the 10 controllers were able to reliable navigate the
robot. This result indicates that the functioning of the evolved CTRNNs was
specific to the sensory-motor conditions encountered in the simulator. This observation
confirms our earlier results that CTRNNs, despite of their ability to
display learning-likes abilities, lack the sensory-motor adaptability found in e.g.
plastic Hebbian synapse networks [2]. However, several techniques for reducing
this “reality gap”-problem, by adding noise at different levels of the simulation,
have been proposed [5][6]. With this in mind the simulator was changed in the
following way: Sensor noise levels were increased from 5% to 10%. In addition,
10% uniform noise was added to the distance traveled by each wheel at each
timestep. Furthermore, the initial conditions of each tested individual changed
in the following way: The starting position was randomized within a 4 by 4 cm
square, and the orientation was randomized within the range forward +/- 15
degrees. With these modifications an incremental evolution lasting 20 generations
was launched, seeded with a population from one of the original runs. This
time the transfer to the real robot was perfect. The best individual of the last
generation was able score a fitness value of 20, i.e. finding the reward-zone in
every trial. No significant behavioral differences compared to the simulation were
observed.
4 Experiment 2: Simple T-Maze with Reward Switching
In order to further investigate the learning-like capabilities of CTRNNs the task
for the robot was now made slightly more complex. In experiment 1 the robots
could rely on the fact that the position of the reward-zone remained fixed during
a whole epoch. Evolved robots were able to explore the whole environment during
trial 1 of each epoch in order to locate this position, but would they also be able
to adapt if the reward position was changed later on within the same epoch.
This turned out not to be the case. When testing the individual analyzed in
section 3.1 by placing the reward-zone to left for 5 trials and then switching
the reward position to the right without resetting the neural network, the robot
would continue to turn left a the T-Junction in the trials after the switch took
place.
A new experiment was now set up in order to check if this lack of adaptivity to
environmental changes taking place later on in an epoch was due to a limitation
in the learning capabilities of the network, or simply given by the fact this
condition was never met during evolution. The evolved robots could simply have
found a minimalistic solution. In this new evolution the duration of each epoch
was increased to 10 trials. The reward-zone position remained fixed in the first
5 trials but was then switched to the other side in the 5 last trials of each epoch.
Each individual was still tested for 4 epochs, 2 with the reward initially to the
left and 2 with the reward initially to the right. The fitness function remained
the same, and since each individual was tested for 40 trials in total the maximum
possible fitness was now 40.
The average result of 10 replications of the experiment is shown in figure 6.
The resulting best fitness in the 10 replications varied between 22 in the worst
case and 38 in the best. In the latter case the best individual did realize that the
reward position had changed in trial 6, and was able to locate the new position.
However in some of the following trials it would still turn the wrong way thus
ending up in a the poison-zone. In order to increase the performance the last bit
an incremental evolutionary approach was now applied. The evolutionary conditions
remained the same, but instead of seeding the evolution with a random
population, it was initially seeded with a population consisting of the best individual
of each of the 200 generations from the best replication of the previous
evolution. Again 10 replications were performed. In most of the runs the fitness
level stayed at 38, and even dropped to 32 in one case (graph not shown). It
seemed that level 38 solution was a local optimum which was difficult to escape.
In one replication, however, the fitness level reached the maximal value of 40, and
when tested afterwards the best individual from this replication could reliably
solve the task. When the reward position switched at trial 6 of each epoch, the
robot would at first move towards the previous location, but when not finding
the reward-zone here anymore it would turn around and initiate a search until
the new reward position was located. In the remaining trials of each epoch the
robot then again turned directly towards the reward-zone, resulting in the total
fitness of 40.
Page 1, 2, 3, 4, 5, 6
Tech Materials (Free)
Robot Behaviors Exploring the T-Maze: Evolving Learning-Like Robot Behaviors using CTRNNs
Humanoid Robotics A Biochemical Subsystem for a Humanoid Robot
Industrial Automation Systems Applying Agents for Engineering of Industrial Automation Systems
Robot Team Cooperation A Descriptive Model of Robot Team and the Dynamic Evolution of Robot Team Cooperation
Kuka Robots For ONU ONU Robotics Technology Center of Excellence, powered by KUKA Robotics Corporation
Augmented reality Annotation System for Robotic Application
Modular Robots Self-Reconfiguration Planning Of Identical Modules
Autonomous robots A New Approach To Robotics
Robotic Mounting Flat Panel Displays With Robotic Mounting
Calibration of Industrial Robots A Photogrammetric Robot Calibration System Based On Off-The-Shelf Low Cost Hardware Components
More... Amazon Books
|
