liuchaoyi/encoder_two_tubes_02_bread_02

Tactile encoder: two_tubes_02 + bread_02

JAX/Flax tactile ResNet18 encoder trained by the train_encoder project in FRS_Tact. The frozen image target encoder is openai/clip-vit-base-patch16.

Training data

• KaiyueChen/two_tubes_02
• KaiyueChen/bread_02

Both datasets were converted locally to LeRobot v3.0. The model pairs each wrist RGB camera with its two tactile cameras and learns against future RGB CLIP embeddings.

Result

• Training stopped early at epoch 12.
• Best checkpoint: epoch 6 (best/).
• Best validation Recall@1: 0.954874.
• Best validation Recall@5: 0.999797.
• Left validation Recall@1: 0.949915.
• Right validation Recall@1: 0.959834.
• Validation uses a sampled same-side retrieval pool of 32 candidates.

last/ contains the epoch-12 state. history.csv and training_curves.png contain the complete run history. The exact training configuration is stored as training_config.yaml.

Loading

Use train_encoder.utils.checkpoint.load_tactile_encoder from FRS_Tact and point it at the downloaded best/ directory.