DART-PT-Flow-LIBERO
ODEWorld image-goal PT-Flow checkpoint retrained with DART (Decode-Augmented Rollout Training) on official LIBERO demonstrations, starting from ldxxx/ODEWorld-PT-Flow-LIBERO. DART keeps the representation frozen (DINOv2 encoder, RAE decoder, goal predictor) and retrains only the flow with decode-path supervision, repairing the motion collapse of the pretrained and latent-only-retrained flows. The checkpoint includes the frozen DINOv2 backbone.
Training configuration
- Window 5, decode-path supervision weight 0.5, DINOv2 region loss weight 0.5, 4 clusters
- 26,000 steps, batch size 2, learning rate 1e-5, seed 20260910
- Selected by lowest validation loss over the task-balanced 130-task Scale-M split
Results (offline open-loop, 64-step image-goal rollouts at 256×256)
| Protocol | Metric | Latent-only parent | DART (this checkpoint) |
|---|---|---|---|
| 100-demo full protocol | mean / median L1 ↓ | 10.311 / 9.64 | 9.944 / 9.30 |
| hard-5 subset | motion concentration conc3 ↓ (GT 0.099) | 0.236–0.447 | 0.145 |
| hard-5 subset | total motion ↑ | 21–58 | 69 |
| 650-demo split | PSNR ↑ / LPIPS ↓ | 19.24–19.45 / 0.196–0.212 | 19.67–19.77 / 0.190–0.195 |
Usage
Load with the release code:
from models import Dinov2PTflowImgoal
flow = Dinov2PTflowImgoal.from_pretrained("pzr114514/DART-PT-Flow-LIBERO")
Training, evaluation, and inference scripts: github.com/Mangkhut160/icassp2027
Scope
This checkpoint is trained and evaluated purely as an offline open-loop predictor on LIBERO demonstration frames. It is not a robot policy: no LIBERO/MuJoCo environment is started, no actions are executed, and no task success rate is measured.