SceneActBench: Can Agents Act on the 3D Scenes They See?
Vision-language model (VLM) agents are increasingly being used to interact with 3D environments rather than simply describing them. However, current benchmarks often focus on static tasks, such as answering questions about a scene or performing simple, single-object edits. This paper introduces SceneActBench, a new benchmark designed to evaluate how well AI agents can perform complex, multi-object actions within 3D environments. By requiring agents to act on complete scenes based on visual input, the benchmark provides a more rigorous test of an agent's true 3D spatial and kinematic understanding.
A Unified Testing Environment
To ensure a fair comparison across different AI models, the researchers developed a standardized evaluation harness. In this setup, an agent is given visual evidence—such as PNG images or sampled video frames—and must interact with a 3D environment using a shared tool interface. The agent operates in a "headless" mode, meaning it uses code to manipulate 3D assets in Blender without a graphical interface. The agent’s final output is then compared against hidden ground truth data using specific geometric metrics to determine how accurately the agent reconstructed or modified the scene.
Five Core 3D Tasks
SceneActBench evaluates agents across five distinct tasks, each testing a different aspect of 3D reasoning:
Layout: Testing spatial grounding by requiring the agent to map 2D images to the correct 3D positions and rotations of furniture.
Camera: Testing egocentric spatial reasoning by asking the agent to infer the 3D position and orientation of a camera based on a single view of a scene.
Articulated: Testing kinematic reasoning by having the agent identify movable parts and reproduce their motion from video frames.
Reconstruction: Testing shape imagination by requiring the agent to build and texture 3D meshes from partial views.
Dynamic: Testing dynamic reasoning by asking the agent to reconstruct the motion of multiple objects over time.
Performance and Limitations
The researchers tested eleven proprietary VLM configurations using this benchmark. The results showed that overall performance varied significantly, with scores ranging from 38.6 to 50.2. Notably, no single model performed consistently well across all five tasks, suggesting that different models have distinct strengths and weaknesses in their 3D capabilities. By analyzing where and how these models failed, the study provides a diagnostic look at the current limitations of VLM agents, highlighting that while these models are becoming more capable, they still struggle to maintain accuracy when tasked with complex, multi-object 3D coordination.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!