This is needed for training the embedding with diverse subtasks. The state/action space of all tasks must be merged or the same (tricky).
Initial set of subtasks:
Pick
Place
Pick with obstacles
Place with obstacles
Stack a block on another block
Servo to a location
two possible options:
implement a parameterized version of the SawyerEnvs which switches the reward function based on a parameter(s) and use the existing MultiTaskEnv (which takes a single environment and a dictionary of args/kwargs to initialize it with)
modify MultiTaskEnv to support different environment classes
This is needed for training the embedding with diverse subtasks. The state/action space of all tasks must be merged or the same (tricky).
Initial set of subtasks:
two possible options: