Skip to content

Question about proprio information in training data #65

Open
@Jianyi2004

Description

@Jianyi2004

Question about proprio information in training data

Question

I have two questions about the training data:

  1. Does the training data include relative distance information between the robot arm and objects?

  2. Could you clarify what information is contained in the proprio array? From the code, I can see it's constructed from state_obs_keys in OXE_DATASET_CONFIGS, but I'm not sure about its specific contents.

Context

I'm looking at the data loading code, particularly:

  • prismatic/vla/datasets/rlds/dataset.py: where proprio is constructed from state_obs_keys
  • prismatic/vla/datasets/rlds/oxe/materialize.py: where dataset configs are loaded

The code shows that proprio is loaded when load_proprio=True, but I couldn't find documentation about what specific robot state information it contains.

Additional Information

  • I've enabled load_proprio=True in my configuration
  • I've modified the data collator to include robot_states in the batch
  • I'm trying to understand what state information is available during training

Would appreciate any clarification on this. Thanks!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions