Skip to content

Latest commit

 

History

History
25 lines (22 loc) · 1.21 KB

patch_notes.md

File metadata and controls

25 lines (22 loc) · 1.21 KB

Patch Notes

Dirty laundry

  • Color jitter is manually implemented for ACT in _robomimic_to_act_data rather than using the ObsUtils color jitter

  • hardcoded extrinsics in val_utils.py

  • Added ac_key under base Algo in robomimic, I suppose this could just access the model.global_config

  • I haven't tested whether aloha_to_robomimic_v2 works with highlevelGMMPretrain

  • Remember: type == 0 is robot, type==1 is hand

  • Now that our hdf5's have either hand or robot data, we can just specify this from the config. So simply set config.train.data_type and config.train.data2_type to hand or robot.

  • Make sure to set train.prestacked_actions = True when dataset actions are prestacked. Hand data is prestacked and of shape (N, T, 3) bc coordinate frame changes each step. Set seq_length to load should be 1 for this case

Dataloader should output batch with following format. Not currently using dones or rewards

dict with keys:  dict_keys(['actions', 'rewards', 'dones', 'pad_mask', 'obs', 'type'])
actions: (1, 30)
rewards: (1, 1)
dones: (1, 1)
pad_mask: (1, 1)
obs: dict with keys:  dict_keys(['ee_pose', 'front_img_1', 'pad_mask'])
        ee_pose: (1, 3)
        front_img_1: (1, 480, 640, 3)
        pad_mask: (1, 1)
type: int