Wrong assumptions about board_shape in toy envs
timorl opened this issue · comments
@alok brought this up in slack, in some places in the code we assume that board_shape has length 3, while for the toy envs it has length 2. There are three offending lines, just grep board_shape[2].
The problem seemed to be caused by a change in the shape of the observations when adding the option to have transitions in the obervations of the ai-safety-gridworlds. This was adapted accordingly in the observations the toy environments return, but not in the observation space which is used in safe-grid-agents to determine the shape of observations initially. For the fix, see david-lindner/safe-grid-gym#28