Restructuring Ideas #279
Description
Objective:
- Simplify code
- Clearer interfaces and responsibilities of components
- Support other geometric primitives for reprojection e.g. individual footholds
- Support more supervision signals - e.g: binary, velocity tracking, locomotion reward, physical properties
- Support for supervision on images (current), top-down view maps
Main - Sugggestion:
- Splitting the learning node into a supervision signal node and a learning node
- Learning node keeps the two graphs but now for the proprioceptive graph we directly replace it with a geometry primitivas graph - messages are sent to the learning node containing geometries and multiple supervision signals extracted by the supervision signal node - this can be interpreted as adding more responsibility to the ANYmal Converter Node.
- Providing a clearer interface of how to implement different neural networks
- Reworking the configuration such that all ros configurations are also part of the ExperimentParam
To be discussed - If we implement it like this everything strongly depends on ros which is also not so desirable for all the message passing - is there a way to joing the best of both worlds :/