Staircase policy speeds robot action planning with streaming inference
Staircase Policy: Streaming Inference for World-Action Models with Large Action Chunks
RoboticsArtificial IntelligenceComputer Vision and Pattern Recognition
Summary
Robots plan actions better when they predict what might happen next, but this prediction takes extra time and slows things down. The authors created a new method called the Staircase Policy that breaks big action plans into smaller pieces and updates these plans step-by-step using the latest observations. This lets robots start acting quickly while still improving their future moves, making long sequences of actions more efficient without needing to replan everything from scratch. Their approach made robots perform tasks faster and more accurately in tests.
What this means in practice
- •For robotics engineers: Increase robot task execution speed and accuracy by using the Staircase Policy to manage long action sequences with fewer computation delays.
- •For real-time system developers: Deploy streaming inference methods to improve responsiveness and throughput in systems requiring continuous action updates.
Authors
Guoheng Sun, Chen Chen, Jin Wang, Ang Li, Teresa Lv
Abstract
World-Action Models (WAMs) improve robotic manipulation by conditioning action generation on predicted future observations, but future prediction adds further inference overhead to already expensive iterative action generation. Action chunking can amortize this cost over multiple actions, yet performance degrades over long execution horizons because later actions remain conditioned on stale observations. We introduce STAIRCASE POLICY, a streaming inference and training framework that turns a flow-matching VLA into a JEPA-style WAM and partitions a large action chunk into sub-chunks at staggered denoising stages. Near-term actions are executed as soon as they become available, while later actions continue to be refined. At each sub-chunk boundary, the future latent is re-predicted from the latest observation and used to update all unexecuted actions, enabling long-horizon execution without repeated full policy inference. The resulting future-prediction error can further serve as a signal for adaptive chunking. S-WAM achieves 97.7% on LIBERO and 87.9% on LIBERO-Plus, and improves performance across multiple policy backbones and real-robot tasks. It reaches 292.7 executed actions per second, $3.62\times$ the throughput of conventional execution at comparable accuracy, while reducing time-to-first-action from 123.6 to 73.3 ms. With additional inference optimizations, throughput further increases to 642.9 actions per second.