Instructions to use jesbu1/robometer-4b-fft-busybox-tiled with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use jesbu1/robometer-4b-fft-busybox-tiled with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, RBM processor = AutoProcessor.from_pretrained("jesbu1/robometer-4b-fft-busybox-tiled") model = RBM.from_pretrained("jesbu1/robometer-4b-fft-busybox-tiled", device_map="auto") - Notebooks
- Google Colab
- Kaggle
jesbu1/robometer-4b-fft-busybox-tiled
Robometer (RBM) reward model, full fine-tune of robometer/Robometer-4B on
tiled multi-view video from AutoRLBench SO-101 push_green_button rollouts.
Tiled input specification
Three synchronized 426x240 h264 views are tiled into one 640x540 RGB frame:
top-> primary upper panel, 640x360front-> lower-left cell, 320x180wrist-> lower-right cell, 320x180
Every panel keeps the source ~1.78 aspect ratio. Tiling utility:
robometer/utils/tiling.py:tile_synchronized_views.
Feeding a different layout still returns numbers but invalidates the rewards.
Training data
jesbu1/so101_busybox_push_green_button_tiled (converted from
AutoRLBench/so101_busybox_push_green_button_cell03_rollouts_426x240).
- 2,900 episodes, single task:
push the green button - Labels: 1,450
successful, 1,450failure(timeouts) - Successes are terminal-flagged from frame 0; timeouts are all-zero
- Source: pi05 DSRL/RLT rollouts plus offline ACT/SmolVLA/Diffusion/DiT/GMM evals
partial_successis intentionally unset: training treats failures as rejected trajectories against same-task successes
Metrics (1,000 steps, 4x H200, lr 2e-5, bs 16)
Training-time evaluation on the full 2,900-episode dataset:
| Step | Pearson | Kendall | Ranking acc |
|---|---|---|---|
| 250 | 0.798 | 1.000 | 1.000 |
| 500 | 0.836 | 1.000 | 1.000 |
| 750 | 0.844 | 1.000 | 1.000 |
| 1000 | 0.844 | 1.000 | 1.000 |
Standalone evaluation of this checkpoint (200 sampled trajectories, frame steps):
| Metric | Value |
|---|---|
| Reward-alignment Pearson | 0.974 |
| Policy-ranking Kendall | 0.996 |
| Ranking accuracy | 0.998 |
Kendall saturates at 1.0 because success vs timeout is nearly perfectly separated on this task; Pearson on graded progress is the informative metric.
Serving
This checkpoint accepts tiled frames directly, or raw per-view videos that the eval server tiles on the fly:
sbatch run_scripts/serve_tiled_eval.sbatch /path/to/checkpoint 8010
python scripts/send_videos_to_tiled_eval.py --url http://localhost:8010 \
--task "push the green button" \
--view top=top.mp4 --view front=front.mp4 --view wrist=wrist.mp4
See robometer/evals/README_TILED_SERVER.md.
Citation
If you use this model, please cite Robometer.
- Downloads last month
- 52
Model tree for jesbu1/robometer-4b-fft-busybox-tiled
Base model
Qwen/Qwen3-VL-4B-Instruct