Retrieve-then-Steer: Online Success Memory for Test-Time Adaptation of Generative VLAs
Jianchao Zhao, Huoren Yang, Hu Yusong, Yuyang Gao, Qiguan Ou, Cong Wan, SongLin Dong, Zhiheng Ma, Yihong Gong
THE PROBLEM
This paper focuses on Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions.. This paper solves the critical problem of frozen Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. models degrading in real deployments by letting robots build a memory of their own successful actions during Simulation & Sim-to-RealDeploymentPutting the trained system on a real robot. and reuse them at test time—no retraining needed. The retrieve-then-steer mechanism retrieves relevant successful Core ConceptsActionA command the robot sends to its motors, controller, or low-level system. chunks and blends them into the Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions.'s Core ConceptsActionA command the robot sends to its motors, controller, or low-level system. generation, improving Data, Distributions & Training IssuesTask successWhether the robot completed the task correctly. rates especially on long-horizon tasks without parameter updates. Read the paper by tracking the Core ConceptsTaskThe job the robot is supposed to complete, such as pick-and-place, navigation, or drawer opening. definition, the Core ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. or data assumptions, and the evidence that supports the claimed improvement.
HOW IT WORKS
Task framing
Core method
Data and supervision
Evaluation evidence
KEY RESULTS
This paper solves the critical problem of frozen Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. models degrading in real deployments by letting robots build a memory of their own successful actions during Simulation & Sim-to-RealDeploymentPutting the trained system on a real robot. and reuse them at test time—no retraining needed. The retrieve-then-steer mechanism retrieves relevant successful Core ConceptsActionA command the robot sends to its motors, controller, or low-level system. chunks and blends them into the Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions.'s Core ConceptsActionA command the robot sends to its motors, controller, or low-level system. generation, improving Data, Distributions & Training IssuesTask successWhether the robot completed the task correctly. rates especially on long-horizon tasks without parameter updates.
WHY DEVELOPERS SHOULD CARE
This paper solves the critical problem of frozen Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. models degrading in real deployments by letting robots build a memory of their own successful actions during Simulation & Sim-to-RealDeploymentPutting the trained system on a real robot. and reuse them at test time—no retraining needed. The retrieve-then-steer mechanism retrieves relevant successful Core ConceptsActionA command the robot sends to its motors, controller, or low-level system. chunks and blends them into the Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions.'s Core ConceptsActionA command the robot sends to its motors, controller, or low-level system. generation, improving Data, Distributions & Training IssuesTask successWhether the robot completed the task correctly. rates especially on long-horizon tasks without parameter updates.
LIMITATIONS
The main limitation to check is whether the claimed behavior holds outside the paper's reported setup. That means testing across different Core ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. embodiments, scenes, objects, and data distributions.
WHAT COMES NEXT
The practical next step is independent reproduction with clear baselines, ablations, and stress tests. For a developer, the useful follow-up is to map the paper's Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. assumptions onto a concrete Core ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. stack, then test the smallest version of the method that could run end to end.