SparkVLA: Stop-Aware Hierarchical VLA with Adaptive Action Chunking for Long-Horizon Manipulation
2026-08-17 • Robotics
Robotics
AI summaryⓘ
The authors address a problem in Vision-Language-Action systems where decisions about when to stop a task and how long to perform an action are connected but usually handled separately. They created SparkVLA, which treats stopping and action length choices as one combined ranking task, improving decision-making without needing extra tuning. Their approach uses a special encoding method to remember task goals and past actions, helping the system focus on important visual details. Tests on a robotics benchmark show that SparkVLA performs much better than previous methods, and real-robot experiments confirm these improvements work outside of simulations.
Vision-Language-Action systemshierarchical decision makingaction chunkingtask terminationself-attentionordinal rankingvisual-token pruningRoboCerebrarobotic task executionsubtask boundary
Authors
Xunyao Lei, Renjun Wu, Tianlin Huo, Xuesong Li
Abstract
At every re-observation point in a hierarchical Vision-Language-Action (VLA) system, two interface decisions must be made: when to terminate the current subtask and how far to execute the proposed action chunk. These decisions are mutually dependent---the optimal stopping point depends on what the executor plans to do, while the optimal execution length depends on where the subtask boundary lies---yet existing architectures evaluate them in isolation, an asymmetry neither module can overcome alone. We present SparkVLA, a stop-aware hierarchical VLA that resolves this mutual dependency by formulating both decisions as a single ranking: Stop competes against every action-prefix length in a unified candidate set, and the system selects the highest-scoring option, eliminating threshold tuning and requiring only offline ordinal preferences. An Anchor-Conditioned Context Encoding module caches a history-aware subtask anchor encoding onset-state memory and goal semantics, guiding visual-token pruning toward task-relevant regions; a Stop-Aware Action-Prefix Selection head scores all candidates via full self bnattention at chunk boundaries for efficiency. On RoboCerebra, SparkVLA achieves 47.12% success rate, surpassing the official hierarchical baseline by 30.57% and the strongest reproducible method by 26.83% Real-robot experiments on multi-step tasks further validate these gains on physical hardware.