---
language:
- en
library_name: lerobot
base_model: nvidia/GR00T-N1.7-3B
datasets:
- Spa-Bench/spa-bench-training-teleoperation-1200
pipeline_tag: robotics
tags:
- robotics
- vision-language-action
- so-101
- spa-bench
license: apache-2.0
---

# GR00T-N1.7 Full Fine-Tune — Spa-Bench Epoch 12

This checkpoint is released as an anonymous supplementary artifact for a paper under double-blind review. It was evaluated on a physical SO-101 robot in Spa-Bench.

## Model details

| Field | Value |
| --- | --- |
| Checkpoint | End of epoch 12; step 76,596 |
| Inputs | Middle RGB, wrist RGB, six joint positions, and a text instruction |
| Outputs | Six absolute joint-position targets |
| Action horizon | 16 |
| Adaptation | Language, visual, projector, VLLN, and diffusion-action modules updated |
| Optimizer | AdamW; learning rate 1e-5; weight decay 1e-5 |

Training-data documentation: [`spa-bench-training-teleoperation-1200`](https://huggingface.co/datasets/Spa-Bench/spa-bench-training-teleoperation-1200).

## Physical evaluation

25/120 familiar trials succeeded. Evaluation stopped before the withheld-composition protocol, so this is a partial result. These are physical rollout results, not simulation metrics.

Rollouts: [`spa-bench-eval-rollouts-groot-n1-7-full-finetune-partial`](https://huggingface.co/datasets/Spa-Bench/spa-bench-eval-rollouts-groot-n1-7-full-finetune-partial).

## Limitations and safety

Training used a two-camera projection of the 1,200-episode release. This condition received a small upward initialization assist, limiting direct ablation claims. Robot policies can move hardware unexpectedly. Use conservative limits, an accessible emergency stop, a clear workspace, and direct supervision. Do not deploy unattended or in safety-critical settings.

## Double-blind release note

Author, institution, source-repository, and archival citation details are intentionally omitted during review. They will be restored in the archival release.