Can Test-Time Scaling Unlock More Reliable Vision‑Language‑Action Robots?

Background

Vision‑Language‑Action (VLA) models have shown strong performance in visual‑motor control, yet their robustness in complex real‑world scenarios remains limited.

Key Finding – Test‑Time Scaling Law

The authors discover that increasing the amount of computation at inference by generating many candidate actions and verifying them (a generate‑and‑verify paradigm) systematically improves generalization and reliability of VLA models. Across several VLA architectures (CogACT, Octo, OpenVLA, SpatialVLA) the action error follows a power‑law relationship with the number of sampled actions.

Method Overview

Stage 1 – Training an Action Verifier

Using a robot dataset, the VLA samples N candidate actions per state, which are clustered into K representative actions. The difference between each candidate and the ground‑truth action (RMSE) is used to synthesize a preference dataset. A vision‑language model is fine‑tuned as an action verifier with a Bradley‑Terry loss augmented by a preference‑strength term.

Stage 2 – Test‑Time Scaling During Inference

At deployment, the system samples \hat N initial actions, fits a Gaussian to their translation and rotation components, and uses majority voting for gripper state, yielding \hat K refined proposals. The trained verifier ranks these proposals and selects the best action for execution.

Experimental Results

RoboMonkey, the proposed framework, improves performance on several benchmarks:

+25 % success on out‑of‑distribution real‑world tasks.

+9 % on the in‑distribution SIMPLER environment.

+7 % on the LIBERO‑Long benchmark.

Additional benefits include reduced grasping errors, fewer task‑progress failures, and fewer collisions.

Scaling Synthetic Data for the Verifier

Increasing the size of the synthetic preference dataset yields a near‑log‑linear improvement in verifier accuracy, which translates into higher success rates in the SIMPLER environment.

Efficient Deployment

The authors implement a dedicated VLA serving engine on top of SGLang that supports rapid repeated sampling and Gaussian perturbation, lowering inference overhead. With higher‑bandwidth memory (HBM), the same latency budget allows greater throughput, further enhancing VLA generalization.

Conclusion

The work establishes a test‑time scaling law for embodied VLA models, provides a scalable verifier‑training pipeline, validates the effectiveness of generate‑and‑verify at inference, and demonstrates practical deployment strategies that keep latency low while boosting performance.