Representation¶

Everything that has to be stored not only temporarily in BOLeRo is located in bolero.representation.

Behaviors¶

The purpose of this library is learning behaviors. Behaviors typically implement the interface Behavior. It is important to know that there are some conventions regarding the inputs and outputs of the generic behavior interface:

Often get_output and set_input deal with the same kind of data, e.g. a DMPBehavior generates a set of positions, velocities and accelerations in Cartesian space or in joint angle space and then they expect to get back the same information from a sensor that measures the actual positions, velocities and accelerations. However, that does not have to be the case for other kind of behaviors.
If a behavior generates multiple derivatives of the same attribute, e.g. positions, velocities and accelerations, it is not per se clear which layout would be to best to put these information in a flat vector. We agreed that all positions (p), all velocities (v) and all accelerations (a) should be stored contiguously, e.g. when the behavior controls 3 joints, the output would have the layout pppvvvaaa. The reason is that it is easy to extract e.g. the position vector with a slice from the output vector.

The following table gives an overview of the behaviors that are provided by BOLeRo.

Behavior name	Usecase	Inputs	Outputs	Parameters
Constant	anything	anything	constant	none
Random	anything	anything	random	none
Linear	anything	anything	linear combination of the inputs	weights of the linear mapping
Dummy	function optimization	parameters	will be used as output	allows direct optimization of the objective function
Dynamical Movement Primitive	trajectories in joint space or Cartesian space	positions, velocities, accelerations	positions, velocities, accelerations	weights of the internal function approximator
Sequence of DMPs	trajectories in joint space	positions, velocities, accelerations	positions, velocities, accelerations	weights of the internal function approximators and subgoals

Constant Behavior¶

A ConstantBehavior always produces a constant output that cannot be changed. It can be used as a behavior baseline.

Random Behavior¶

A RandomBehavior always produces a random output that is completely random and normal distributed. It can be used as a behavior baseline.

Linear Behavior¶

A LinearBehavior generates a linear mapping \(y = W x\) from an input vector \(x\) (with an additional bias component that is always 1) to an output vector \(y\).

Dummy Behavior¶

A DummyBehavior always produces the output that has been given as parameters from the optimizer. It can be used in cases where no behavior is required actually, e.g. for plain function optimization or where the behavior is encoded in the environment.

Dynamical Movement Primitive¶

Dynamical movement primitives represent trajectories DMPBehavior, e.g. in joint space. They can generalize over several meta-parameters (goal, velocity at the goal, execution time) and can be learned from demonstrations. A variant of DMPs that works in Cartesian space is CartesianDMPBehavior.