Xingxin on Bug Different datasets and controllers define the state space and action space of a robotic gripper differently. Here is a quick cheat sheet comparing gripper state and action spaces across RoboCasa, CRISP, and DROID.
RoboCasa
Here is the convention for RoboCasa.
State
The state tracks the position of each finger independently:
gripper_qpos[0]: left finger jointgripper_qpos[1]: right finger joint
If your gripper’s maximum width is 0.08 m, then the fully open positions are:
gripper_qpos[0]:0.08 / 2 = 0.04gripper_qpos[1]:-(0.08 / 2) = -0.04
Action
The action is a discrete boolean value.
- open:
-1 - close:
+1
CRISP
Here is the convention for CRISP - Compliant ROS2 Controllers for Learning-Based Manipulation Policies.
State
The state is a single scalar value within the range [0, width]. If your gripper’s maximum width is 0.08 m, the state space is [0, 0.08].
Action
The action is a boolean value.
- open:
1 - close:
0
DROID
Here is the convention for the DROID. It uses a Robotiq 2F-85 gripper and normalizes the values.
State
The state is a continuous, normalized value in the range [0, 1]:
- fully open:
0 - fully close:
1
Action
Unlike RoboCasa and CRISP, the action space is a continuous value in the range [0, 1]:
- fully open:
0 - fully close:
1