Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

RoboMonkey: Scaling Test-Time Sampling and Verification for Vision-Language-Action Models

Created by
  • Haebom

Author

Jacky Kwok, Christopher Agia, Rohan Sinha, Matt Foutter, Shulu Li, Ion Stoica, Azalia Mirhoseini, Marco Pavone

Outline

This paper presents a test-time scaling technique to improve the robustness of the Vision-Language-Action (VLA) model in unstructured real-world environments. We study how to improve the robustness and generalization performance of VLA through sampling and validation, and show that the relationship between the action error and the number of generated samples follows an exponential power law. Based on this, we propose RoboMonkey, a test-time scaling framework for VLA. RoboMonkey generates multiple action samples from VLA, adds Gaussian noise, generates an action proposal distribution through majority voting, and then selects the optimal action using a VLM-based verifier. We train a VLM-based action verifier through a synthetic data generation pipeline, and demonstrate the performance improvement of VLA using RoboMonkey through simulations and hardware experiments. The experimental results show an absolute performance improvement of 25% on out-of-distribution tasks and 9% on in-distribution tasks, and show that fine-tuning the VLA and action verifier together improves the performance by 7% compared to fine-tuning the VLA alone when adapting to a new robot setting.

Takeaways, Limitations

•
Takeaways:
◦
We present RoboMonkey, an effective test time scaling technique that improves the robustness and generalization performance of VLA models.
◦
Achieve performance improvements on both out-of-distribution and in-distribution tasks through a sampling and validation-based approach.
◦
Demonstrating the utility of learning VLM-based action verifiers using synthetic data.
◦
A method for simultaneous fine-tuning of VLA and action verifier to improve adaptability to new robot configurations is presented.
•
Limitations:
◦
RoboMonkey's performance improvements may be limited to specific experimental environments and VLA models.
◦
Further research is needed on the generalization performance of synthetic data generation pipelines.
◦
Further validation of RoboMonkey's robustness in complex and diverse real-world environments is needed.
◦
The computational cost of VLM-based verifiers may increase.
👍
Made with Slashpage