CLAP Adapts VLMs to VLAs with Language-Action Grounding.

Yuri Ishitoya, Jeremy Siburian, Masashi Hamaya, Kuniaki Saito, Cristian C. Beltran-Hernandez, Mai Nishimura· July 13, 2026 View original

Summary

CLAP (Causal Language-Action Prediction) is a method that directly converts pretrained Vision-Language Models (VLMs) into Vision-Language-Action models (VLAs) with minimal architectural changes. It achieves this by prepending natural-language action descriptions to numeric action sequences, causally conditioning action prediction on a language-action plan, and significantly improving performance and robustness with single-epoch fine-tuning.

This research introduces CLAP (Causal Language-Action Prediction), a novel approach for directly adapting powerful Vision-Language Models (VLMs) into Vision-Language-Action models (VLAs) with minimal architectural modifications. The primary challenge in this adaptation is the mismatch between the VLM's pretrained language distribution and the numerical action sequences required for robotic control. CLAP addresses this by conditioning the prediction of precise action tokens on a natural-language action plan. The method involves prepending each numeric action sequence with a corresponding natural language description, effectively guiding the VLM's generation towards the desired actions while preserving its core capabilities. With just a single epoch of fine-tuning, CLAP demonstrates significant performance gains, achieving 90.8% on the LIBERO benchmark and enhancing robustness against various perturbations. The researchers plan to release CLAP as an open-weight, multi-scale VLA family, facilitating further analysis of VLM-to-VLA capability transfer.

Why it matters

This method offers a more transparent and efficient way to leverage the semantic understanding of large VLMs for robotic control, accelerating the development of capable and robust embodied AI systems.

How to implement this in your domain

  1. 1Explore CLAP for adapting existing VLMs to robotic control tasks with minimal fine-tuning.
  2. 2Integrate language-action planning into your robot learning pipelines to improve action prediction.
  3. 3Utilize the open-weight CLAP models to benchmark and develop new VLA applications.
  4. 4Investigate how causal language conditioning can enhance other multimodal AI systems.

Who benefits

RoboticsManufacturingLogisticsAutonomous VehiclesHealthcare

Key takeaways

  • CLAP directly converts VLMs to VLAs with minimal architectural changes.
  • It uses language-action grounding to bridge the VLM-VLA output distribution gap.
  • Single-epoch fine-tuning yields significant performance and robustness improvements.
  • CLAP offers a transparent path to understanding VLM capability transfer to robotics.

Original post by Yuri Ishitoya, Jeremy Siburian, Masashi Hamaya, Kuniaki Saito, Cristian C. Beltran-Hernandez, Mai Nishimura

"arXiv:2607.08974v1 Announce Type: cross Abstract: Vision-language-action models (VLAs) inherit semantic capabilities from pretrained VLMs, yet large-scale post-training on robot data and architectural modifications can reshape the backbone so extensively that it becomes difficult…"

View on X

Originally posted by Yuri Ishitoya, Jeremy Siburian, Masashi Hamaya, Kuniaki Saito, Cristian C. Beltran-Hernandez, Mai Nishimura on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

Data-Efficient Deep Learning Guidelines for Inertial Sensor Classification

This study provides empirical guidelines for estimating the minimum training set size needed for deep learning models in inertial sensor classification tasks. It reveals that accuracy follows a consistent logarithmic growth pattern, allowing for data-efficient planning of recording campaigns and achieving practical stability with fewer samples than traditionally assumed.

Ofir Kruzel, Itzik KlienJul 13, 2026
AI Engineering & DevToolsAI Research

On-Device Adaptive AI Boosts EV Battery Power Prediction

Researchers developed a novel approach for on-device learning in electric vehicles (EVs) that continuously adapts pretrained battery power prediction models to new data. This method significantly improves forecasting performance, reducing mean absolute errors by up to 14.88% with offline adaptation and 7.49% with online adaptation.

Avik Bhatnagar, Anton Paule, Tobias Schuermann, Sebastian Reiter, Oliver BringmannJul 13, 2026
AI ResearchAI Engineering & DevTools

New Differentiable Logic Networks Outperform Fixed-Connection Models

Researchers introduce a novel method for optimizing connections in deep differentiable logic gate networks (LGNs) and lookup table networks (LUTNs), achieving superior performance with significantly fewer gates. The approach allows for parallel learning of optimal gate types and LUT entries, demonstrating improved accuracy on benchmarks like MNIST.

Wout Mommen, Lars Keuninckx, Matthias Hartmann, Werner Van Leekwijck, Piet WambacqJul 13, 2026