Loading dataset...
CachedTrajectoryDataset: 256 demos, 8192 samples
  Source: {'cache_root': '/data/libero/ood_objpos_circle_overlay', 'benchmark': 'libero_spatial', 'task_ids': [0], 'frame_stride': 3}
  Total: 8192 samples
✓ Train: 7783 samples
✓ Val: 409 samples

Initializing model (type=para)...
Loading DINOv2 model...
✓ DINOv2 backbone is trainable
✓ DINO embedding dim: 384
✓ Learnable start keypoint embedding (dim=384)
✓ Feature convs: 3× Conv2d(3×3) at pred_size=64
✓ Volume   head → (B, 4, 32, 64, 64)
✓ Gripper  head → (B, 4, 2, 64, 64)  [1×1 conv, CE 2-class]
✓ Rotation head → (B, 4, 3, 32, 64, 64)  [1×1 conv, CE]
Trainable parameters: 32,880,520 / 32,880,520 (100.00%)

Computing dataset stats from random subset: 500/8192 samples (seed=42)
/data2/cameron/miniconda3/envs/uva/lib/python3.10/site-packages/wandb/sdk/data_types/image.py:324: DeprecationWarning: 'mode' parameter is deprecated and will be removed in Pillow 13 (2026-10-15)
✓ Saved stats cache: /data/cameron/para_normalized_losses/libero/checkpoints/para_pretrain_3d_circle/dataset_stats.json
✓ Height range from dataset: [0.917348, 1.177162] m
✓ Gripper range from dataset: [-1.000000, 1.000000]
✓ Rotation range (delta rotvec): ['-0.093', '-0.092', '-0.002'] .. ['0.040', '0.209', '0.025']
✓ Reference rotation: ['0.9996', '0.0014', '-0.0284', '-0.0013']
✓ Position range from dataset: ['-0.413', '-0.366', '0.917'] .. ['0.125', '0.273', '1.177']

Starting training for 9999 epochs...
✓ Rotation loss SKIPPED
✓ EMA loss weights initialized (vol=11.8, rot=3.5, grip=0.69)
  self._image = pil_image.fromarray(
Epochs:   0%|          | 1/9999 [2:10:12<21697:55:18, 7812.81s/it]

============================================================
Epoch 0/9999
============================================================
                                                                                                              

  [async eval] logged video to wandb: ep000_fail.h264.h264.h264.h264.mp4

  [async eval] logged video to wandb: ep000_fail.h264.h264.h264.h264.h264.h264.h264.mp4
Train Loss: 0.5420 (Volume: 0.2788, Gripper: 0.2633, Rotation: 0.0000)
Val - Loss: 0.3029, Volume: 3.1433, Pixel Error: 9.49px, Height Error: 14.981mm, Gripper: 0.1247
✓ Saved best model (val_loss=0.3029)

⏱ Time limit reached (130.2 / 20 min). Stopping.