Comparison of joint trajectories. The DiffusionPolicy approach shows discontinuities at action‑chunk boundaries, while our method yields trajectories with higher continuity.
Comparison of joint trajectories. The DiffusionPolicy approach shows discontinuities at action‑chunk boundaries, while our method yields trajectories with higher continuity.
Dynamic Tasks (Successful & Failed)
Robot Camera View, At Original Speed (Ours)
Robot Camera View (Raw Action)
Pick & Place, At Original Speed
Acceleration comparison between raw actions and our proposed method. The raw action representation produces large accelerations at the boundaries of action chunks, which leads to jitter. In contrast, our method reduces boundary spikes and improves smoothness.
Comparison of action-chunk boundary smoothing methods: our B‑spline refitting versus the weighted fusion. The y‑axis shows acceleration magnitude, and the x‑axis represents the smoothing time window.
Key hyperparameters used in our model.
To ensure consistent testing conditions, we placed the manipulated objects at the locations specified in the table. For example, for the stack-block task, we positioned the red block for picking and the gray block for placing according to the table, and conducted evaluations at these points.
Comparison of joint trajectories. The blue line represents the executed trajectory, the green line is the model's predicted output, and the red line shows the refined trajectory after applying continuity-constrained refitting.
Comparison of joint trajectories. The blue line represents the executed trajectory, the green line is the model's predicted output, and the red line shows the refined trajectory after applying continuity-constrained refitting.
