Event
R2A Uses Motion Primitive Graph Retrieval to Improve VLA Robustness
1 reports · 1 sources
Pham Tri Quang and other authors propose Retrieve to Act (R2A), a test-time scaling framework that organizes successful demonstrations into an offline spatiotemporal motion primitive graph. At inference time, a query-relevant graph neural network locates the current execution state, retrieves structurally compatible future motion primitives, and then fuses them with the predicted actions of a frozen VLA through per-dimension gating. The authors state that the framework requires no fine-tuning of the base policy. The authors report that R2A achieves up to +12% absolute success rate improvement on the simulation benchmarks LIBERO-Plus and RoboTwin2.0-Plus, and +40% improvement in real-world deployment. These figures are results self-reported by the authors.
Report timeline
- Retrieve to Act: Improving VLA Execution Robustness with Motion Primitive Graph Retrieval
The authors propose Retrieve to Act (R2A), a test-time scaling framework that organizes successful demonstrations into an offline spatiotemporal motion primitive graph. At inference, a query-dependent graph neural network locates the current execution state and retrieves structurally compatible future motion primitives, which are then fused with the frozen VLA's predicted actions through dimension-wise gating.
arXiv cs.RO ·