Reproduce “ELASTIC: Efficiently Learning to Adaptively Scale Test-Time Compute for Generative Control Policies” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Andrew Zou Li, Gokul Swamy, Yonatan Bisk, Andrea Bajcsy
An algorithm called ELASTIC learns to adaptively scale test-time computation for Generative Control Policies (GCPs) in robotics, dynamically adjusting sequential denoising steps and parallel action samples. This approach consistently enhanced task success while reducing inference latency by up to 34% on real-world robot manipulation, outperforming fixed compute baselines.
Key insights
- Adaptive, state-dependent allocation of both sequential denoising steps and parallel action samples is crucial for optimizing GCP performance and efficiency, consistently outperforming static or single-axis scaling.
- The learned meta-policy can autonomously identify and respond to task-specific bottlenecks, prioritizing precision through sequential steps when needed, or exploring multimodal actions via parallel samples.
- The system allows base GCPs to compensate for inherent weaknesses or uncertainties by dynamically allocating more resources in states where the policy is less competent or faces ambiguity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.