Towards Long-horizon Embodied Agents with Tool-Aligned Vision-Language-Action Models

2026-05-13

AI Summary

Existing vision-language-action (VLA) models struggle to simultaneously achieve effective global planning and diverse physical manipulation in long-horizon tasks. This work proposes the VLAs-as-Tools framework, which employs a high-level vision-language model for global planning and recovery, while delegating local subtasks to multiple specialized VLA tools. A unified tool interface enables event-triggered replanning, tightly coupling planning and execution. The approach introduces a novel VLA tool-family interface and a Tool-Aligned Post-Training (TAPT) method to enhance tool-call fidelity and coordination. Evaluated on LIBERO-Long and RoboTwin, the framework improves task success rates by 4.8% and 23.1%, respectively, and increases tool-call fidelity by 15.0%.

Abstract

Vision-language-action (VLA) models are effective robot action executors, but they remain limited on long-horizon tasks due to the dual burden of extended closed-loop planning and diverse physical operations. We therefore propose VLAs-as-Tools, a strategy that distributes this burden across a high-level vision language model (VLM) agent for temporal reasoning and a family of specialized VLA tools for diverse local physical operations. The VLM handles scene analysis, global planning, and recovery, while each VLA tool executes a bounded subtask. To tightly couple agent planning with VLA tool execution in long-horizon tasks, we introduce a VLA tool-family interface that exposes explicit tool selection and in-execution progress feedback, enabling efficient event-triggered agent replanning without continuous agent polling. To obtain diverse specialized VLA tools that faithfully follow agent invocations, we further propose Tool-Aligned Post-Training (TAPT), which constructs invocation-aligned training units for instruction following and adopts tool-family residual adapters for efficient tool specialization. Experiments show that VLAs-as-Tools improves the success rate of $π_{0.5}$ by 4.8 points on LIBERO-Long and 23.1 points on RoboTwin, and further enhances invocation fidelity by 15.0 points as measured by Non-biased Rate. Code will be released.

Problem

• long-horizon tasks
• vision-language-action models
• embodied agents
• tool alignment
• physical operations

Innovation

• Vision-Language-Action Models
• Tool-Aligned Post-Training
• Long-horizon Planning
• Embodied Agents
• Modular Robotics