Kinematic graph retrieval steadies vision-language-action policies under visual shifts

R2A retrieves motion primitives from past demonstrations and fuses them with frozen policy actions, lifting success rates by up to 12 points in simulation and 40 points on a real robot.

Research Lab

Pham Tri Quang · Bao-Ngoc Dao · Minh Duc Nguyen · Duc-Duy Nguyen · Hung T. Ho · Linh Dang Le · +12 more

VinUniversity · VinRobotics · Hanoi University of Science and Technology · German Research Center for Artificial Intelligence (DFKI) · University of Oldenburg

Research Digest·28 October 2026·2 min read

Illustration: Kinematic graph retrieval steadies vision-language-action policies under visual shifts

In one line

R2A makes VLA robot policies robust to visual changes by retrieving geometric motion primitives from training experience.

Before you read this

Three ideas this paper builds on

vision-language-action modelA robot policy model that maps visual observations and language instructions to actions, trained on demonstrations and often used for manipulation tasks.56 papers →test-time scalingAllocating extra computation during inference to improve output quality.86 papers →diffusion modelA type of generative model that learns to reverse a gradual noising process to create data.152 papers →

Read it as

Robots that follow language instructions using vision often fail when lighting, camera angles, or backgrounds change. Their AI over-relies on visual appearance, causing it to lose track of the correct physical movements.

R2A ignores these corrupted visual cues. It maps the robot's current pose onto a graph of successful past movements and retrieves the geometrically correct next motion. This memory is then blended with the robot's own prediction to keep the execution on track.

Tested across three leading robot AI systems, R2A improved success rates by up to 12% in simulation and 40% in real-world scenarios without any retraining. This makes physically deployed robots far more reliable under unpredictable real-world shifts.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

vision-language-action model test-time scaling motion primitive graph diffusion model graph neural network retrieval augmented generation