Potential discrepancy in action execution: Robot moves despite action being set to current ee_state

[Likeooo]

I am currently working on a VLA project and utilizing VLABench for evaluation. I encountered an unexpected behavior when testing a baseline policy.

To verify the control consistency, I modified the RandomPolicy to output the exact same values as the input ee_state (End-Effector state), effectively commanding the robot to remain stationary. However, upon reviewing the video playback of the test run, the robot still exhibits visible movement.

Steps to Reproduce:

1. Use the RandomPolicy within the VLABench environment.
2. Modify the RandomPolicy.predict(self, obs, **kwargs) to set the predicted action equal to the input ee_state . My code is as below:

def predict(self, obs, **kwargs):
        delta_pos = np.random.uniform(-0.1, 0.1, 3)
        delta_euler = np.random.uniform(-0.1, 0.1, 3)
        gripper_open = np.random.uniform(0, 1, 1)
        current_ee_state = obs["ee_state"]
        if len(current_ee_state) == 8:
            pos, quat = current_ee_state[:3], current_ee_state[3:]
            euler = quaternion_to_euler(quat)
        elif len(current_ee_state) == 7:
            pos, euler = current_ee_state[:3], current_ee_state[3:]
            
        # target_pos = np.array(pos) + delta_pos
        # target_euler = euler + delta_euler
        # modify the above two lines to below two lines
        target_pos = np.array(pos)
        target_euler = euler

        gripper_state = np.ones(2)*0.04 if gripper_open >= 0.1 else np.zeros(2)
        return target_pos, target_euler, gripper_state

3. Run the evaluation and observe the video output/execution. My eval script is as below:

from VLABench.evaluation.evaluator import Evaluator
from VLABench.evaluation.model.policy.base import RandomPolicy

demo_tasks = ["select_fruit"]
save_dir = "./logs"

evaluator = Evaluator(
   tasks=demo_tasks,
   n_episodes=2,
   max_substeps=10,   
   save_dir=save_dir,
   visulization=True
)

random_policy = RandomPolicy(model=None)
result = evaluator.evaluate(random_policy)

Observed Behavior:
The robot arm moves/drifts instead of remaining perfectly still as commanded.

Questions:

1. What is the underlying cause of this movement? (e.g., Is it due to controller latency, coordinate system mismatches, IK errors, or physics engine noise in the simulator?)
2. To what extent will this execution error/drift affect the overall benchmark results for VLA models? I am concerned that this discrepancy might lead to an underestimation of model performance.