Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution

Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution

The paper’s authors introduce Retrieve to Act (R2A), a method intended to help Vision-Language-Action (VLA) policies execute more robustly when visual inputs shift outside the conditions they encountered in training. The authors say R2A organizes successful training demonstrations into a motion primitive graph, then uses a Graph Neural Network to retrieve compatible future motions during execution. A learned gating mechanism combines those motions with the frozen VLA’s predicted actions, aiming to correct spatial errors while retaining the policy’s local reactivity. The authors report tests with three VLA architectures and say R2A improved success rates in simulation benchmarks and real-world deployments without finetuning the base policies. This suggests the method may draw on motion patterns learned during training to address some execution errors caused by changed visual context. The abstract does not establish how well the reported results transfer to other robots, tasks, or deployment conditions, or how much the method’s performance depends on the available demonstrations. What conditions and task types account for the reported real-world improvement, and how consistently does it carry over to new settings?

Publication: e6664374-d950-482e-81d4-c4d85bce5f23

attribution — pass: The reading consistently attributes the method and its results to the paper's authors ('The authors introduce', 'The authors say', 'The authors report'), matching the arXiv abstract. No source claim is presented as established fact or as Escape Meridian's view.

We introduce Retrieve to Act (R2A), a test-time scaling framework

summary — uncertain: The summary is faithful, but its lead never says this is about robots. 'Vision-Language-Action (VLA) policies' is the central term and is left unexplained: these are AI models that take camera images and language instructions and output robot actions. 'Robots' first appears only in the limitations sentence. A non-specialist may not grasp what is being made more robust. Smallest repair: add 'robot-control' before 'VLA policies' in the first sentence, or add a tooltip for VLA. Also suggest library tooltips for motion primitive, Graph Neural Network, gating mechanism and frozen. The numbers (up to +12% in simulation, +40% real-world) are omitted but appear in the claims.

Vision-Language-Action (VLA) policies demonstrate remarkable semantic generalization, yet remain highly vulnerable to out-of-distribution (OOD) visual shifts during execution.

date — pass: originalPublished is 2026-10-27. This matches the v1 submission date on the arXiv page. It is not the later feed date of 2026-10-29.

[Submitted on 27 Oct 2026]

support — pass: Each claim is supported by a verbatim excerpt. The single inference ('This suggests the method may draw on motion patterns…') is hedged and consistent with the abstract's framing of kinematic priors from training experience. The limitations and the question correctly note that the abstract lacks task-level and deployment details. They do not invent costs or benefits, and they do not turn benchmark gains into proven real-world reliability.

they fundamentally overlook the rich, uncorrupted kinematic priors embedded in the policy's own training experience

media — pass: The selected media is the verified direct link to the arXiv abstract page for this paper. It is the most direct access to the original source.