Lecture 5 - Evolutionary Robotics Flashcards
What is the goal of evolutionary robotics?
Evolutionary robotics uses artificial evolution to automatically generate control systems and morphologies for autonomous robots.
What are the 5 steps of an evolutionary algorithm?
- Generate a series of genomes which define neural networks.
- Upload the neural networks to the robot and evaluate the fitness of the robot.
- Selection - select the best robot’s genomes
- Crossover and mutation
- Replace the genomes of previous generation, repeat until solution is found.
What happens if you put a evolutionary robot in complex environment with a simple fitness function? (give example)
Sustainability can be evolved.
Robot with a battery of 20s put in an area with a charger in the corner.
Even tho the fitness function didn’t include battery, it evolved to use the charger 2s before dying.
Can you evolve vision? If so how would you do it?
Yes you can.
One example was based on linear vision. 64 photo receptors. Only 16 at the centre used. A Laplacian filter was used to detect areas of contrast and scalled between 0 and 1.
This was fed into a neural network that could take a time series of data.
This neural net learnt to avoid areas of high contrast.
How would you evolve learning?
Instead of evolving hardcoded weights you can evolve leaning rules that govern the weight of the synapse.
What are the 4 learning rules in the lecture? Briefly describe them.
- Hebb - weights on both sides of the neuron are strengthened when they fire.
- Post synaptic - the weight after the synapse is strengthened when they fire
- Pre synaptic - the weight before
- Covarience - combination of the two
What are the important aspects of evolving learning?
- neural networks can use different learning rules in different parts of the system
- There is no need of teacher or reinforcement learning, no gradient descent
and local minima - The system can learn and adapt in new environments - more robust
- Evolved adapted individuals can transfer smoothly from simulated to real world.
What are Framsticks?
Evolutionary evolved creatures made up from sticks. Neurons and sensors are hosted on sticks.
What’s was the Golem project? What was significant about the project?
Similar the the framsticks simulation but the robots are 3D printed. The simulated evolved neural nets were downloaded onto a PIC controller the robot and tested.
The evolution took place in the simulation, but actually tested on robots.
What is it called when evolved Robots don’t work as well as expected in real life?
Reality gap.
What is a disadvantage when trying to make evolved robots in real life?
Often the evolved systems are very complicated and therefore difficult and expensive to build.
What is exploration-exploitation? What problem does it solve?
The co-evolution of robots and the controller.
The reality gap.
Talk me through the algorithm for exploration-exploitation? *(co-evolution)
- Evolve the robots in initial simulator, test the best on a real robot whilst recording sensory signals.
- Evolve the simulators to match sensory signals of evolved simulated robots with those recorded on real robots.
- Repeat until a good match is found.
How many iterations of co-evolution did it take to make a successfull star shaped robot in 2007?
- Proof that it is a good mechanism to cross the reality gap.
If a robot is broken, how can it to recover?
Map of different solutions (walking gaits) is in the controller of the robot. (evaluated by a fitness function) The robot was able to navigate this map to test out different walking gaits until it found a successful gait for its injury. - typically this took less than 40 seconds when tested on the 6 legged robot.
Not so much evolution but a search of feature space.
