Abstract
Left: simulation, Right: Hardware, Center: overlap(green) between hardware(red) and simulation(blue)
One main challenge in mobile robotics is the autonomous locomotion. With modular robots, the number of possible structures is exponentially dependent of the number of robots and the controller solutions may not be engineerable. Coevolution is a way to obtain gaits: in this case we optimize both the structure of the robot and the controller. But performing online optimization with the hardware is really time consuming. Another possibility is to use simulation, which allows optimizing rapidly a solution. The main disadvantage of simulation is the reality gap: a good gait in simulation can be nonadapted to the real environment. One way to make the simulation match the real world is Hybrid Optimization, in which we do not want to optimize only a controller or a structure but the parameters of the simulation. This optimization is composed of repetitions of the three following steps: optimize gaits in simulation, then transfer the gaits to the hardware and finally optimize the simulation to make it match the real world via the metaoptimization. This master project follows several studies on Roombots, especially the work of Yura Perov during summer 2012. During his internship, he performed a first cycle of hybrid optimization using a metamodule, i.e. a structure composed by two modules connected. This master project extends this approach to structures composed by two to four modules.
Achievements
• New hardware record for the highest speed with a simple metamodule: 8.04 cm/s (previous one was:
5.8cm/s) This was done using for the first time hybrid rotation-oscillation CPG network instead of pure oscillation and was only optimized in simulation.
• Increasing the proportion of usable particles for the online optimization from 29% to 75%
• High speed gaits found using a metamodule with compliant element: 11.7 cm/s and 9.2 cm/s depending
on the type of element used (see section 2.3 ).
• Successful transfer from simulation to hardware of gait using other structures than a metamodule:
tripod, snake and star. It is the first time that another hardware structure than a metamodule is
efficient with the hardware
• Second cycle of metaoptimization performed with different structures: the type of ground property
used for locomotion depends on the strategy chosen and for example optimizing the simulator with a
type of gait using crawling will not help to simulate sliding gaits.
• Successful transfer from simulation to hardware using coevolution. Both the structure and the controller
were optimized in simulation and the result on the hardware reaches a speed of 5.96 cm/s.
• Realization of reconfiguration between two structures composed by 3 Roombots modules for the first time