XS-VLA: Teaching Tiny Vision-Language-Action Models with Spatial Supervision and Demonstration Conditioning
XS-VLA: Teaching Tiny Vision-Language-Action Models with Spatial Supervision and Demonstration Conditioning
This paper introduces XS-VLA, a staged training framework for compact Vision-Language-Action models that uses coarse-grained spatial distillation to initialize the backbone and latent flow matching conditioned on demonstrations to learn action-space velocity fields. Evaluated on LIBERO benchmarks, XS-VLA achieves an average success rate of 90.25% across two training seeds, outperforming a SmolVLA-0.25B base model trained under the same settings (86.00%). On three Mobile ALOHA tasks, the method increases average strict success from 21.7% to 65.0%. Ablation studies confirm the utility of auxiliary representation initialization and regularized demonstration-conditioned flow learning for improving control performance in compact VLA policies.
The study presents XS-VLA, a compact Vision-Language-Action model trained via spatial distillation and demonstration conditioning, achieving 90.25% average LIBERO success in two seeds and improving Mobile ALOHA strict success from 21.7% to 65.0% without requiring the teacher model at deployment.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
training seeds
2
View original evidence
en conditions an action-space velocity field on a demonstration latent, using KL regularization while jointly optimizing the backbone and action modules. The deployed policy contains 243.99M parameters and operates without the teacher or posterior encoder. XS-VLA achieves 90.25% average LIBERO success in each of two training seeds, compared with 86.00% for a SmolVLA-0.25B base trained under our settings. Ablations ex
Open source S3Mobile ALOHA tasks
3
View original evidence
amine both training stages through matched image pretraining and Huber/MSE controls. On three Mobile ALOHA tasks, average strict success increases from 21.7% to 65.0%. These results demonstrate the control utility of auxiliary representation initialization and regularized demonstration-conditioned flow learning for compact VLA policies
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
en conditions an action-space velocity field on a demonstration latent, using KL regularization while jointly optimizing the backbone and action modules. The deployed policy contains 243.99M parameters and operates without the teacher or posterior encoder. XS-VLA achieves 90.25% average LIBERO success in each of two training seeds, compared with 86.00% for a SmolVLA-0.25B base trained under our settings. Ablations ex
Open source S3
amine both training stages through matched image pretraining and Huber/MSE controls. On three Mobile ALOHA tasks, average strict success increases from 21.7% to 65.0%. These results demonstrate the control utility of auxiliary representation initialization and regularized demonstration-conditioned flow learning for compact VLA policies
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org