A similar error occurs with "TSP-v1"

Hi, thank you for raising this issue! Will look into this shortly.

Cool!
FYI it's part of this PR
pytorch/rl#2129
which enables rendering of envs using from_pixels=True in torchrl

Sort of personal request: It'd be nice to have a way to ask for headless rendering / rgb_arrays in a consistent way (from my 10K feet high view it seems envs have a different way of dealing with rendering). Something like make(envname, rgb_array=True) or render(rgb_array=True).
For torchrl, we prefer always getting arrays of uint8 and deliver them to the users as part of their env output and leave them deal with them as they want (if the user wants to render smth they can just call plt.imshow(data["pixels"]))
Having the human rendering by default usually introduces some consequent overhead.

The way the above PR deals with this is to render on screen, record the content of the plt image in a buffer, convert that buffer to a tensor (since there is no generic way of getting a rgb).

Sorry it took so long @vmoens but this should be fixed now 😄