Hugging Face's logo

jesbu1
/
robometer-4b-fft-armnet-tiled

Transformers
Safetensors
qwen3_vl
reward_model
rbm
preference_comparisons
Model card Files
xet
 Community

jesbu1/robometer-4b-fft-armnet-tiled

Model Details

  • Base Model: Qwen/Qwen3-VL-4B-Instruct
  • Model Type: qwen3_vl

Tiled Armnet Specification

This checkpoint was full-fine-tuned on synchronized Armnet camera views tiled into one 640 x 540 RGB frame.

  • Single-arm layout: large top view above front and wrist views.
  • Bimanual layout: large top view above left_wrist and right_wrist views.
  • Native episode labels preserved: successful, failure, and suboptimal.
  • Tiling utility: dataset_upload/helpers.py:tile_synchronized_views.
  • Training datasets:
    • jesbu1/armnetbench_v01_tiled_so101
    • jesbu1/armnetbench_v01_tiled_bimanual_so101

Training-time metrics at step 1000:

Dataset Pearson Kendall
Single-arm SO-101 0.9224 0.9933
Bimanual SO-101 0.9553 0.9333

Citation

If you use this model, please cite:

Downloads last month
1
Safetensors
Model size
4B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for jesbu1/robometer-4b-fft-armnet-tiled

Base model

Qwen/Qwen3-VL-4B-Instruct
Finetuned
(386)
this model