arxiv:2604.14125

HiVLA: A Visual-Grounded-Centric Hierarchical Embodied Manipulation System

Published on Apr 15

· Submitted by

Authors:

Abstract

HiVLA presents a hierarchical vision-language-action framework that decouples semantic planning from motor control using a diffusion transformer action expert with cascaded cross-attention for improved robotic manipulation.

AI-generated summary

While end-to-end Vision-Language-Action (VLA) models offer a promising paradigm for robotic manipulation, fine-tuning them on narrow control data often compromises the profound reasoning capabilities inherited from their base Vision-Language Models (VLMs). To resolve this fundamental trade-off, we propose HiVLA, a visual-grounded-centric hierarchical framework that explicitly decouples high-level semantic planning from low-level motor control. In high-level part, a VLM planner first performs task decomposition and visual grounding to generate structured plans, comprising a subtask instruction and a precise target bounding box. Then, to translate this plan into physical actions, we introduce a flow-matching Diffusion Transformer (DiT) action expert in low-level part equipped with a novel cascaded cross-attention mechanism. This design sequentially fuses global context, high-resolution object-centric crops and skill semantics, enabling the DiT to focus purely on robust execution. Our decoupled architecture preserves the VLM's zero-shot reasoning while allowing independent improvement of both components. Extensive experiments in simulation and the real world demonstrate that HiVLA significantly outperforms state-of-the-art end-to-end baselines, particularly excelling in long-horizon skill composition and the fine-grained manipulation of small objects in cluttered scenes.

View arXiv page View PDF Project page Add to collection

Community

Violin-Y

Paper submitter 1 day ago

✨ HiVLA is a hierarchical embodied manipulation agent system that combines visual grounding, semantic planning, and robust action execution for long-horizon and fine-grained robotic manipulation.

avahal

about 14 hours ago

the cascaded cross attention inside the DiT action expert is the clever hinge here, it sequentially fuses global features, high-res object crops grounded by the bbox, and a language embedding that encodes the required skill. that design keeps the low-level policy focused on robust execution while the high-level planner handles the grounding, which seems essential for long-horizon tasks in clutter. i'd be curious to see an ablation where you remove the local crop path or replace the language conditioning with just the global features to quantify each piece's contribution. the emergent error correction where the planner can re propose grounding if the diffusion misses a grasp feels promising but i'd like to see more systematic failure cases. btw, the arxivlens breakdown helped me parse the method details, a solid walkthrough that covers this setup; it even covers how the cascaded attention plays into the conditioning: https://arxivlens.com/PaperView/Details/hivla-a-visual-grounded-centric-hierarchical-embodied-manipulation-system-60-0dabfd0c