Pre-trained model checkpoints are available on Hugging Face.
The tokenizer checkpoint is saved at step 100,000, which showed the best reconstruction performance.
Download: tokenizer/checkpoint_step_100000.pt
| Checkpoint | MAE (m) | RMSE (m) | < 1m (%) | VQ Perplexity |
|---|---|---|---|---|
| Step 100,000 | 1.15 | 2.37 | 72.17% | 159.23 |
| Step 110,000 | 3.61 | 4.66 | 2.59% | 145.19 |
| Step 120,000 | 3.45 | 4.35 | 2.37% | 140.94 |
| Step 130,000 | 3.38 | 4.31 | 2.89% | 141.39 |
Selected: Step 100,000 - Best depth reconstruction accuracy with 72.17% of points within 1m error.
The world model checkpoint is saved at step 50,000 (final checkpoint after full training).
Download: world_model/checkpoint_step_50000.pt
| Checkpoint | Step | Val Acc | Val Loss |
|---|---|---|---|
| checkpoint_0046000.pt | 46,000 | 83.47% | 1.4996 |
| checkpoint_0048000.pt | 48,000 | 80.74% | 1.6320 |
| checkpoint_0050000.pt (final) | 50,000 | 84.32% | 1.4603 |
| checkpoint_latest.pt | 50,000 | same | same |
Selected: Step 50,000 - Final trained model with best validation accuracy of 84.32%.