ELASTIC: Learning Robot Scaling Strategies

ybisk · x · 2026-07-13

This repost discusses two dimensions of test-time scaling for diffusion/flow-based robot policies: - **Sequential scaling**: Increasing the number of denoising steps - **Parallel scaling**: Increasing the number of samples The article notes that both methods improve performance but introduce robot execution latency. Furthermore, it is often difficult to predict in advance which dimension should be prioritized during actual operation. To address this, the authors introduce **ELASTIC**, which uses reinforcement learning to learn this selection strategy.

Related event: ELASTIC Enables Robots to Adaptively Allocate Inference Compute(2 posts)→

Original post →

More from Embodied

Embodied channel →