ELASTIC

Efficiently Learning to Adaptively Scale Test-Time Compute for Generative Control Policies

Andrew Zou Li, Gokul Swamy, Yonatan Bisk, Andrea Bajcsy

Carnegie Mellon University

Conference on Robot Learning (CoRL) 2026

Overview

Diffusion and flow policies can spend extra test-time compute in two ways: more denoising steps (sequential) or more candidate actions (parallel). Which one helps depends on the task, the state, and the policy.

ELASTIC learns this with reinforcement learning. For a frozen policy, it learns a state-dependent schedule that sets sequential depth and parallel width together, in one decision per denoising step. On a real robot, it matches Best-of-10 success with 34% less latency.

Method

A meta-policy watches the candidate actions as they denoise and, at each step, chooses how far to advance each one, or prunes it. It is trained on task success minus a compute cost.

Allocating during denoising

Each marker is a candidate. As they denoise, red is pruned, green is slowly denoised, and yellow advances with a larger step size. Then verifier picks the best finished candidate.

The right amount of samples to keep, and the right step size to advance each, is not obvious. ELASTIC learns to adapt this based on the environment state and the denoising samples themselves to maximize task success minus a compute cost.

Denoised action trajectory rollout at the current denoising step

Insufficient compute along both axes lead to different failure modes

Too few steps can lead to imprecise grasps or placing failures, while too few samples can result in sampling a spurious mode and reaching for the wrong mug.

Results

Diffusion Policy

Each orange point is a meta-policy trained at a different compute cost. At matched compute it beats Fixed and Sequential-only, and often approaches Best-of-N with far fewer steps.

Six panels (PushT, PushT Obstacle, Square PH, Square MH, Can Paired, Can Reverse) plotting success or coverage against sequential steps. ELASTIC's orange frontier lies above or left of budget-matched fixed and sequential-only baselines, approaching the Best-of-N reference line at far fewer sequential steps.

π₀.₅ VLA on LIBERO-10

Matches Sequential-Only Scaling with 6% less latency than the base policy.

Success rate against inference latency on LIBERO. ELASTIC is the lowest-latency point at about 94 percent success, above budget-matched fixed and the base policy, while Best-of-N reaches slightly higher success at roughly 1.5 times the latency.

Heatmaps over end-effector position for the ten LIBERO-10 tasks, in two rows of five. For each task, the upper tile shows sequential steps and the lower tile shows parallel samples; both tend to rise near grasp and place locations, and tasks differ in which axis they use. — Where compute goes across the ten LIBERO-10 tasks

π₀.₅ VLA on Real-World Mug Task

Matches Best-of-10 success with 34% less latency.

Success rate against inference latency on a real Franka arm. ELASTIC reaches the 80 percent Best-of-N success level at about 62 milliseconds, whereas Best-of-N needs about 96 milliseconds; fixed schedules and the base policy sit far lower.

Browse rollouts

Policy Task Episode

Execution

Allocation Timeline

BibTeX

@inproceedings{li2026elastic,
  title     = {{ELASTIC}: Efficiently Learning to Adaptively Scale
               Test-Time Compute for Generative Control Policies},
  author    = {Li, Andrew Zou and Swamy, Gokul and Bisk, Yonatan and Bajcsy, Andrea},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}