language:
- en
library_name: transformers
base_model: Qwen/Qwen2.5-VL-3B-Instruct
datasets:
- Spa-Bench/spa-bench-training-teleoperation-1200
pipeline_tag: robotics
tags:
- robotics
- vision-language-action
- so-101
- spa-bench
license: other
license_name: qwen-research
license_link: https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/blob/main/LICENSE
VLA-0 — 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; 77,246 optimizer updates |
| Inputs | Middle RGB, wrist RGB, six joint positions, and a text instruction |
| Action representation | Six joint targets encoded as text with 1,000 bins |
| Action horizon | 8 |
| Adaptation | Full Qwen backbone; no LoRA or QLoRA |
| Optimizer | AdamW; learning rate 5e-6; weight decay 0.01 |
Training-data documentation: spa-bench-training-teleoperation-1200.
Physical evaluation
0/72 familiar physical rollouts 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-vla-0-partial.
Limitations and safety
Training used a motion-trimmed projection of the 1,200-episode release. Deployment requires the VLA-0 inference stack; this is not a drop-in LeRobot policy. 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.