How to set stateencoding and actionencoding in another robot (legged robot)

[wuyukun-tong]

Hi
I am very confused hoe to modify the following code to fit my own data

`class StateEncoding(IntEnum):
"""Defines supported proprio state encoding schemes for different datasets."""

NONE = -1  # no state provided
POS_EULER = 1  # EEF XYZ + roll-pitch-yaw + 1 x pad + gripper open/close  
POS_QUAT = 2  # EEF XYZ + quaternion + gripper open/close
JOINT = 3  # 7 x joint angles (padding added if fewer) + gripper open/close
JOINT_BIMANUAL = 4  # 2 x [6 x joint angles + gripper open/close]

class ActionEncoding(IntEnum):
"""Defines supported action encoding schemes for different datasets."""

EEF_POS = 1  # EEF delta XYZ + roll-pitch-yaw + gripper open/close
JOINT_POS = 2  # 7 x joint delta position + gripper open/close
JOINT_POS_BIMANUAL = 3  # 2 x [6 x joint pos + gripper]`

In our custom dataset the robot state are the imu data and joint position data and the action is velocity of robot.

[kpertsch]

If you want to do finetuning you don't need to actually modify this code, you can just directly modify the finetuning config: https://github.com/octo-models/octo/blob/main/scripts/configs/finetune_config.py

The StateEncoding and ActionEncoding is primarily used to determine the action normalization mask (ie which dimensions should be normalized) here:

octo/octo/data/oxe/__init__.py

Line 37 in 5eaa5c6

if dataset_kwargs["action_encoding"] is ActionEncoding.EEF_POS:

So if you want to use your dataset for OpenX co-training, you can define a new element in the Enums above and then modify the code in the link above to set the normalization mask accordingly (eg for Octo we typically don't normalize the gripper dimension)

[wuyukun-tong]

Thank a lot !