rllab-postech/grasp-left-wrist-finetune-bg2-left-pretrain-ae256

Grasp Left Wrist Fine-tune from BG2 Left Pretrain AE256

Final fine-tuned policy for the rllab-postech/grasp target dataset.

Setup

• —Fine-tune dataset: rllab-postech/grasp
• —Pretrain checkpoint: rllab-postech/bg2-left-wrist-latent-flow-pretrain-ae256
• —Camera: observation.images.cam_wrist_left
• —State transform: bg2_left_arm (19D -> 8D)
• —Action transform: bg2_left_arm_link_xyz (19D -> 39D FK points)
• —Vision encoder: DINOv3, frozen
• —Action autoencoder: point_temporal
• —Latent action dim: 256
• —Horizon: 12
• —Fine-tune steps: 50000 total, resumed from the 10000-step baseline

Result

• —10000-step baseline val loss: 0.160404
• —Best observed val loss: 0.128479 at step 33000 (not checkpointed)
• —Best saved checkpoint val loss: 0.137849 at step 40000
• —Final 50000-step val loss: 0.139152
• —Final train loss: 0.109001

Recommended checkpoint: checkpoints/grasp_left_wrist_from_bg2_left_pretrain_40000step_best_saved.pt.

Files

• —checkpoints/grasp_left_wrist_from_bg2_left_pretrain_40000step_best_saved.pt
• —checkpoints/grasp_left_wrist_from_bg2_left_pretrain_50000step_final.pt
• —config.yaml
• —metadata.json
• —normalizer.json
• —results/run_summary.json
• —results/finetune_50k_resume.log