Robots Learn to Allocate Inference Compute on Demand
CSProfKGD · x · 2026-07-13
This repost introduces a robotics study: ELASTIC teaches robots "when to think harder and when to think less."
The core idea is that diffusion/flow-based robot policies have two test-time scaling methods:
- Sequential scaling: increasing denoising steps;
- Parallel scaling: increasing the number of samples.
Both improve performance but increase robot latency, so actions shouldn't be treated equally. The paper uses RL to learn how to allocate compute across different contexts: e.g., using fewer steps for free-space movement, and more samples during grasping to explore different grip modes.
Related event: ELASTIC Enables Robots to Adaptively Allocate Inference Compute(2 posts)→
More from Embodied
- PROWL uses a world model to keep Minecraft agents exploring after failures — nathanbenaich · 2026-07-21
- LeRobot v0.6.0 adds end-to-end 3D depth training data for robots — RemiCadene · 2026-07-21
- NVIDIA brings its Cosmos 3 Edge world model to Jetson for on-device robot control — liu_mingyu · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21
- DIY wooden box packs 6 Intel Arc Pro B70 cards with a FreeCAD model — nick_ziv · 2026-07-21
- Snake-like robot moves on fully passive wheels and winding motion — ___Mufasaa · 2026-07-21