SMILE

We compare SMILE with raw-action SmolVLA to test whether predicting B-spline coefficients improves real-robot task success. No official code is released, so we reimplemented SMILE’s action interface from the paper on SmolVLA’s pretrained backbone. The action expert denoises 8 cubic B-spline control points that are decoded into the action chunk. Both models predict relative translation and rotation-vector actions with absolute gripper widths and use matched training settings within each task. This round re-evaluates the first round’s models under blended execution without retraining them.

• Prediction length. Both models predict 50 action steps, following the SmolVLA recipe. This prediction length is not confirmed in the SMILE paper.
• Episode-end supervision. We fit splines only to valid samples and supervise only the coefficient directions they determine, filling unconstrained directions with the dataset mean. This retains partial windows in both arms; full windows use the paper’s coefficient-space loss.
• Fine-tuning budget. Both arms use batch 64, learning rate 1e-4 and at least 20k steps or three epochs, whichever is greater, without image augmentation or a validation split. Exact parity cannot be established because the paper does not report a complete training recipe.
• Blended execution. Both models were trained to execute 15 of their 50 actions per plan, 0.5 seconds at 30 Hz, for synchronous fixed-horizon execution. Here the station requests a new plan once less than 800 ms of actions remain and crossfades plans over 600 ms. Because each plan is shorter than that window, the station replans about every 90 ms and executes about three actions of each plan. Both arms use the same settings, but this replaces the paper’s fixed execution horizon.