(a) Discrete basins of attraction for a two DoF system. Different colors indicate different kinds of attractors, showing a rich variety of behaviors even in a low-dimensional system. (b) Hexapod simulated robot with 18 DoF. The robot is initialized in a random environment and thanks to the repelling potential is able to execute several behaviors in succession. The sweeping through the behavioral landscape is shown in the velocity space (c).
One of the primary questions of this project is understanding how to generate structured behaviors for a robotic system, without specifying any reward nor objective function. The systems we are interested in should be able to exploit the embodiment among brain, body and environment to self-explore a wide range of behaviors and automatically extract a suitable controller for each of them. In order to bootstrap this goal-free explorative process, we use a biologically plausible synaptic mechanism for self-organizing controllers, in particular, differential extrinsic plasticity (DEP), which has proven to enable embodied agents to self-organize their individual sensorimotor development and generate highly coordinated behaviors during their interaction with the environment.
We use a dynamical systems framework to describe a behavior as an attractor in the brain-body-environment system using DEP. The behaviors self-organize within a few seconds of live interaction and are specific to the embodiment of the robot. Each behavior corresponds to a potentially useful motion primitive.
The behavioral landscape generated by DEP is then explored thanks to a "repelling potential" which allows the system to actively explore all its attractor behaviors in a systematic way. With a view to a self-determined exploration of goal-free behaviors, our framework enables switching between different motion patterns in an autonomous and sequential fashion. Our algorithm is able to recover all the attractor behaviors in a toy system and it is also effective in two simulated environments. A spherical robot discovers all its major rolling modes and a hexapod robot learns to locomote in 50 different ways in 30min.
[ ]