Action Segmentation on ALOHA Robot (test)
28.93Edit DistanceHand-Only
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Hand-OnlyTraining Protocol=Zero-shot baseline (Hand-pretrained model direct to robot data), Backbone=X3D-M, Temporal Backbone=MS-TCN++ 14L, Random Seed=4562026.06 | 28.93 | 19.01 | 14.83 | 7 | 13.96 | |
| Robot-OnlyTraining Protocol=Trained from random initialization on 16 robot demonstrations, Backbone=X3D-M, Temporal Backbone=MS-TCN++ 14L, Random Seed=4562026.06 | 70.5 | 43.69 | 36.92 | 18.46 | 40.22 | |
| Human-to-Robot + VLMTraining Protocol=Pre-trained on human source data and fine-tuned on 16 robot demonstrations, Post-processing=Edit-safe VLM correction (InternVL2-8B), Backbone=X3D-M, Temporal Backbone=MS-TCN++ 14L, Random Seed=4562026.06 | 80.68 | 51.97 | 41.81 | 22.98 | 46.42 | |
| Human-to-RobotTraining Protocol=Pre-trained on human source data and fine-tuned on 16 robot demonstrations, Backbone=X3D-M, Temporal Backbone=MS-TCN++ 14L, Random Seed=4562026.06 | 80.7 | 51.23 | 41.36 | 22.22 | 45.21 |