CALVIN environment differences

[Infinity4B]

I want to reproduce the results of qwbu/RoboDual-OpenVLA-Generalist, which is based on this repo. The environment initialization is like https://github.com/OpenDriveLab/RoboDual/blob/0fd80c67c8be3dadf91040cfc52ef6caa00c87a0/vla-scripts/evaluate_calvin.py#L121. I compare the enviroments provided by RoboDual and Dexbotic-Benchmark, and find out that given the same action, the robot moves in different ways. I want to ask that is Dexbotic-Benchmark have some different action spaces against other environments?

[Infinity4B]

I find out that it is probably due to the config setting use_delta. When I set it to false, it works well. It seems that all the models use delta action, but the use_delta here affects this:

        if self.use_delta:
            # Delta mode: add predicted action to previous action
            original_action = np.copy(last_action)
            original_action[6:] = 0  # Reset gripper action
            action = original_action + predict_action
        else:
            # Absolute mode: use predicted action directly
            action = predict_action

I think it doesn't have the need to use original action. What does the setting works with Dexbotic models?