CLAP: Direct VLM-to-VLA Adaptation via Language-Action Grounding
CLAP: Direct VLM-to-VLA Adaptation via Language-Action Grounding
This research presents CLAP (Causal Language-Action Prediction), a method that adapts pretrained vision-language models into vision-language-action models with minimal architectural changes. By prepending numeric action sequences with natural language descriptions, CLAP addresses output-distribution mismatch and preserves VLM capabilities. The open-weight models are released in 0.8B, 2B, and 4B parameter sizes. On the LIBERO benchmark, the 2B model achieves 90.8% success (+14.9 points over VLA-0), and on a real UR5e robot, it reaches 60% success against 15% for a matched baseline.
The paper introduces CLAP, an open-weight family of vision-language-action models that adapts pretrained VLMs to robotic control by grounding numeric action tokens in natural language descriptions, achieving a 14.9 percentage point improvement on the LIBERO benchmark and 60% success rate on a real UR5e robot compared to 15% for a matched baseline.
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.
- paper: not_reported
- code: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ence with a natural-language action description, causally conditioning precise action-token prediction on a language-action plan without modifying the backbone architecture. With only 4-5 hours of training, 2B CLAP achieves 90.8% on LIBERO (+14.9 pt over VLA-0) and improves robustness on LIBERO-PRO under language, object, and spatial perturbations. Ablations attribute the gains to action-grounded language semantics,
Open source S4
and on a real UR5e robot 2B CLAP reaches 60% success against 15% for matched VLA-0. We release CLAP at 0.8B, 2B, and 4B as an open-weight, multi-scale compact VLA family from a single VLM lineage, enabling controlled analysis of VLM-to-VLA capability transfer.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org