| # Expected Inputs | |
| SGA-GSN 3D checkpoints expect the public SGA-GSN preprocessing pipeline to provide: | |
| - Visual/object point input compatible with the AdaPoinTr shape model. | |
| - Tactile point input with `xyz + contact` channels. | |
| - The corresponding object id metadata used by the dataset loader. | |
| For downstream RL, the perception stack uses: | |
| ```text | |
| shape_checkpoint = checkpoints/shape_completion/ap_ps55.pth | |
| grasp_checkpoint = checkpoints/sga_gsn/ppct_bce_vtencdrop_8dep_vds_featex_best.pth | |
| ``` | |
| The 2D CNN and CNNMCA baselines expect the 2D visual/tactile image tensors produced by the SGA-GSN 2D VTG dataset loaders. | |