Is there a form of low dimensional input available?

[lumen2023]

Is there a form of low dimensional input available?

[greenphoenixegg]

Similarly, could you kindly provide improvements for trajectory inputs other than image input?

[NM512]

Hi,

Thank you for your comments. For low-dimensional or state-based inputs, I recommend checking how the dmc_proprio environment is implemented and configured.

You can refer to the configuration here as an example of how observation keys are set up for image and state encoder indivisually:

dreamerv3-torch/configs.yaml

Lines 96 to 103 in 7433d1e

	dmc_proprio:
	steps: 5e5
	action_repeat: 2
	envs: 4
	train_ratio: 512
	video_pred_log: false
	encoder: {mlp_keys: '.*', cnn_keys: '$^'}
	decoder: {mlp_keys: '.*', cnn_keys: '$^'}

This should give you a good reference for handling non-image inputs.