| FlowerVLAPolicy-compatible packaged checkpoint. | |
| Source: Stage-2 joint training (text + optional vision LoRA on Florence-2, | |
| co-trained with action loss + pixel grounding regularizer). | |
| LoRA adapters have been merged into base Linear weights; head dropped. | |
| Source ckpt: /usr/itetnas04/data-scratch-01/yuchwang/data/workspace/robot_learning/Lerobot_flower/checkpoints/stage2_task3_all_v4_v2_1/20260518-123814_job2017180/best | |
| Stage 1 config: {'head_hidden': 256, 'lora_rank': 8, 'lora_alpha': 4, 'lora_dropout': 0.05, 'vision_lora_rank': 0, 'vision_lora_alpha': 4, 'vision_lora_dropout': 0.0, 'vision_lora_stages': []} | |
| Stage 2 config: {'lambda_pixel': 0.0, 'freeze_vlm_stack': True, 'proprio_dropout': 0.0} | |
| Best val (from training): { | |
| "val_loss": 0.06790262050926685, | |
| "step": 17500, | |
| "experiment": "stage2_task3_all_v4_v2_1" | |
| } | |