GFM Cubes 22b
Gaze Flow Matching policy trained on the Blueberry cubes-to-baskets task.
Architecture
- Policy: Gaze Flow Matching (DiT + Q-Former)
- Backbone: ResNet18
- Gaze mode: Heatmap (Dim 2.2) with spatial attention modulation
- Cross-camera: FPV → left/right gaze projection
- Horizon: 16, sampling steps: 8
- Dim model: 512
Training
| Parameter | Value |
|---|---|
| Dataset | fecasado/Ncubes-to-Nbaskets-320x240 |
| Steps | 100,000 |
| Batch size | 128 |
| Optimizer | AdamW (lr=1e-4, betas=0.95/0.999) |
Usage
from lerobot_policy_gaze_fm import GazeFMPolicy
policy = GazeFMPolicy.from_pretrained("fecasado/gfm-cubes-22b")
actions = policy.predict_action_chunk(observation_batch)
- Downloads last month
- 7