TMLR Rebuttal

Improvements of RNR-DP on Response-Sensitive Tasks Compared to Diffusion Policy

Our Responsive Noise-Relaying Diffusion Policy (RNR-DP) consistently outperforms Diffusion Policy (DP) across responsiveness-sensitive tasks. On this page, we qualitatively showcase the advantages of our method in terms of responsiveness and adaptability to environmental changes.

Key Observations:

Pushing a stationary ball on a table into the target area requires a high level of responsiveness. Once the ball loses contact with the end-effector, it continues moving in the direction of its initial velocity, making the task highly sensitive to even slight deviations in action distribution, which can significantly reduce the success rate. Since our method executes one action at a time, it demonstrates greater responsiveness compared to Diffusion Policy and successfully completes the task under the same conditions.

Key Observations:

Relocating a ball on a desk to the goal region using a dexterous hand is particularly challenging due to the complex dynamics of hand-object interactions, especially when relative displacement may occur during the grasping process. A common failure case of Diffusion Policy is its inability to lift the ball off the table within the maximum episode time due to the ball slipping on the surface or shifting within the fingers. In contrast, under the same initial conditions, our method not only successfully achieves a stable grasp but also smoothly transports the ball to the goal region with minimal delay, demonstrating the advantages of RNR-DP.

Key Observations:

When Diffusion Policy predicts a sequence of future actions at a given moment, it executes all predicted actions without adjustment. However, if the predicted motion deviates significantly from the target point (red dot), it lacks the ability to correct itself based on the latest observations, ultimately failing to push the chair to the target area. In contrast, our method continuously updates decisions based on the most recent observations, ensuring high precision in guiding the chair toward the target and completing the task efficiently.

Key Observations:

Pushing the T-shaped block precisely into the target region with a stick is particularly challenging, as it requires fine control over both position and orientation to prevent unintended rotations or misalignments. Due to its long action horizon, Diffusion Policy often introduces slight misalignments, causing the end-effector to oscillate near the edge in an attempt to correct the placement. In contrast, our method, which employs single-action rollouts, accurately positions the object with precision, ensuring correct alignment.