Hi, thank you for open-sourcing this project. I am recently studying the RoboCOIN/Cobot_Magic_fold_clothes dataset.
I have a question about the design of {dir}_gripper_open_scale, which is currently described as:
Value range [0, 1]; 0 means fully closed, 1 means fully open; converted from collected data.
I am not fully convinced this is the most reasonable representation for gripper state/action across different robots.
The main concern is that different grippers can have very different physical characteristics, including: different maximum opening widths, different mechanical structures. Under this setting, learning a normalized “open scale” in [0, 1] may align the semantic meaning of closed ↔ open, but it may not preserve the physically meaningful notion of gripper opening across embodiments. For example, two grippers with the same open_scale = 0.5 may correspond to very different absolute jaw openings in the real world.
This may introduce ambiguity for cross-embodiment learning: the same normalized value does not necessarily correspond to the same physical opening distance, policies may learn embodiment-specific scale conventions instead of physically grounded gripper behavior, transfer between different gripper morphologies may therefore become harder.
I feel that a more meaningful standardized representation would be to use a unified absolute opening distance (for example, in meters), rather than a per-gripper normalized open scale. An absolute measure may better preserve the physical meaning of the gripper state and make cross-robot learning more consistent.
By the way, could you clarify the exact physical meaning and unit of left_gripper_open in observation.state? I initially interpreted it as the absolute jaw opening distance, but that seems unlikely, as a value of 0.5 would correspond to 0.5 m, which is unrealistic for the AgileX gripper.
Thanks.
Hi, thank you for open-sourcing this project. I am recently studying the RoboCOIN/Cobot_Magic_fold_clothes dataset.
I have a question about the design of {dir}_gripper_open_scale, which is currently described as:
I am not fully convinced this is the most reasonable representation for gripper state/action across different robots.
The main concern is that different grippers can have very different physical characteristics, including: different maximum opening widths, different mechanical structures. Under this setting, learning a normalized “open scale” in [0, 1] may align the semantic meaning of closed ↔ open, but it may not preserve the physically meaningful notion of gripper opening across embodiments. For example, two grippers with the same open_scale = 0.5 may correspond to very different absolute jaw openings in the real world.
This may introduce ambiguity for cross-embodiment learning: the same normalized value does not necessarily correspond to the same physical opening distance, policies may learn embodiment-specific scale conventions instead of physically grounded gripper behavior, transfer between different gripper morphologies may therefore become harder.
I feel that a more meaningful standardized representation would be to use a unified absolute opening distance (for example, in meters), rather than a per-gripper normalized open scale. An absolute measure may better preserve the physical meaning of the gripper state and make cross-robot learning more consistent.
By the way, could you clarify the exact physical meaning and unit of left_gripper_open in observation.state? I initially interpreted it as the absolute jaw opening distance, but that seems unlikely, as a value of 0.5 would correspond to 0.5 m, which is unrealistic for the AgileX gripper.
Thanks.