Spatial Representation Distillation and Knowledge Routing for Vision-Language-Action Models
Spatial Representation Distillation and Knowledge Routing for Vision-Language-Action Models
This paper introduces SPARK-VLA, a framework designed to enhance vision-language-action models (VLAs) with three-dimensional spatial understanding without relying on external geometric inputs. The method employs two main components: spatial representation distillation, which transfers 3D knowledge from a frozen expert to a LoRA-adapted visual branch while maintaining the original visual pathway, and a lightweight knowledge router that directs action-relevant visual tokens to the language model based on gradients from action loss. Evaluated on benchmarks including LIBERO, LIBERO-Plus, CALVIN, and real-world manipulation tasks, SPARK-VLA demonstrates strong performance using only a 0.5B parameter language backbone and approximately 53M trainable parameters. The code is publicly available.
The authors propose SPARK-VLA, a framework that distills spatial understanding from a frozen expert and routes visual tokens to a language model without external 3D inputs, evaluated across LIBERO, LIBERO-Plus, CALVIN, and real-world manipulation tasks.
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.
- code: Available Artifact link Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
, and (2) a lightweight knowledge router, which selectively forwards action-relevant visual tokens to the language model using routing supervision derived from action-loss gradients. Across LIBERO, LIBERO-Plus, CALVIN, and real-world manipulation, SPARK-VLA achieves strong performance with only a 0.5B language backbone and approximately 53M trainable parameters, providing an effective and parameter-efficient approach
Open source S4
to 3D-aware VLA learning. Code is available at https://github.com/park-jungin/spark-vla.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org