README.md

Supplementary Videos

This directory contains supplementary videos for the anonymous submission.

S1: TCP4D Tracking

S1_TCP4D_Tracking/ contains examples of TCP tracking. Each example includes camera-view videos and corresponding point-cloud visualizations. The pcd video shows the predicted point cloud.

S2: Augmented Data

S2_augumented_data/ contains example videos organized by source: DROID/, RoboTwin/, and Robolab/.

• original.mp4 and source_demo_aloha-agilex.mp4 show source demonstrations.
• Other filenames identify target robot embodiments, including Franka Panda, Jaco, UR5 with WSG, ARX, and Piper. Files marked cross_embodiment show cross-embodiment examples.

All videos were processed using our cross-embodiment pipeline.

S3: Real-World Task Demonstrations

S3_real_task_demo/ contains videos for two real-world tasks. Filenames beginning with task1 and task2 correspond to the respective tasks; files with the _long suffix show longer versions.

• task1_fast.mp4 and task1_long.mp4: placing pens
• task2_fast.mp4 and task2_long.mp4: stacking bowls

Task 1 shows rollouts for placing pens, while Task 2 shows rollouts for stacking bowls.

TCP4D infer 32 action step for a chunk, and replan when execute 20 steps.

The _fast variant uses the cuRobo library with CUDA acceleration for inverse kinematics (IK) planning. For safety, we adjusted the action execution frequency, so its theoretical speed can be higher than shown.

The _long variant includes a joint movement penalty in the optimization objective to reduce large changes in joint positions during execution. It is implemented using SciPy without CUDA acceleration.

S4: Few_shot_Adaptation

These two sets of videos both demonstrate the effect of transfer learning using ten-shot training. The videos are all unseen, only for evaluation..