SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning

2026-06-11

AI Summary

Current vision-language models (VLMs) are constrained by rigid action interfaces that hinder their ability to balance flexibility and accuracy in open-ended, complex 3D/4D spatial reasoning tasks. This work proposes SpatialClaw, a training-free framework that pioneers code as a stateful, interactive action interface. By leveraging a preloaded Python kernel containing input frames, VLM agents dynamically generate and execute composable perception-geometry operation units over multiple steps, adapting responsively to task demands. This approach overcomes the limitations of single-step execution or fixed tool invocation, achieving an average accuracy of 59.9% across 20 spatial reasoning benchmarks—surpassing prior methods by 11.2 percentage points. Consistent performance gains are observed across six diverse VLM backbones, all without task-specific fine-tuning or model adaptation.

Abstract

Spatial reasoning, the ability to determine where objects are, how they relate, and how they move in 3D, remains a fundamental challenge for vision-language models (VLMs). Tool-augmented agents attempt to address this by augmenting VLMs with specialist perception modules, yet their effectiveness is bounded by the action interface through which those tools are invoked. In this work, we study how the design of this interface shapes the agent's capacity for open-ended spatial reasoning. Existing spatial agents either employ single-pass code execution, which commits to a full analysis strategy before any intermediate result is observed, or rely on a structured tool-call interface that often offers less flexibility for freely composing operations or tailoring the analysis to each task. Both designs offer limited flexibility for open-ended, complex 3D/4D spatial reasoning. We therefore propose SpatialClaw, a training-free framework for spatial reasoning that adopts code as the action interface. SpatialClaw maintains a stateful Python kernel pre-loaded with input frames and a suite of perception and geometry primitives, letting a VLM-backed agent write one executable cell per step conditioned on all prior outputs, enabling the agent to flexibly compose and manipulate perception results and adapt its analysis to both intermediate text and visual observations and the demands of each problem. Evaluated across 20 spatial reasoning benchmarks spanning a broad range of static and dynamic 3D/4D spatial reasoning tasks, SpatialClaw achieves 59.9% average accuracy, outperforming the recent spatial agent by +11.2 points, with consistent gains across six VLM backbones from two model families without any benchmark- or model-specific adaptation.

Problem

• spatial reasoning
• vision-language models
• action interface
• 3D/4D reasoning
• tool-augmented agents

Innovation

• spatial reasoning
• code-as-interface
• tool-augmented agents
• vision-language models
• stateful execution