New 3D AI Model Achieves Spatial Reasoning with Fewer Computing Resources

Spatial understanding and 3D reasoning represent critical capabilities for the next generation of artificial intelligence systems. However, existing approaches that combine vision and language models for these tasks demand substantial computational resources, limiting their practical deployment.

A More Efficient Approach to 3D Understanding#

Researchers have unveiled 3D-MoE, a novel 3D vision-language model that tackles this efficiency challenge through a mixture-of-experts architecture. Rather than activating every component of a neural network for every task, this design selectively engages only the most relevant experts for each input. The system incorporates a specialized routing mechanism that takes into account both the type of data being processed and its spatial context, guided by a unique training approach they call a routing curriculum that stabilizes performance during development.

Extending to Robotic Action#

To enable real-world robotic applications, the team integrated a diffusion-based component called Pose-DiT that generates precise 6D pose actions—capturing position and orientation in three-dimensional space. This component leverages rectified flow, a framework allowing the system to produce accurate pose predictions in a single step rather than requiring multiple iterations. The resulting 3D vision-language-action (VLA) model bridges the gap between visual understanding and physical robot manipulation.

Performance Gains in Testing#

Across multiple benchmarks for 3D vision-language tasks, 3D-MoE demonstrated superior results while using substantially fewer activated parameters than traditional approaches. The model achieves this efficiency without sacrificing capability: it maintains higher success rates for robot manipulation operations and enables real-time inference—a critical requirement for practical robotic deployment where latency directly impacts task success.