Rollback-Free Stable Brick Structures Generation

2026-05-07

AI Summary

This work addresses the challenge that autoregressive models face in generating physically stable block structures while simultaneously satisfying gravity and connectivity constraints. Existing approaches rely on post-hoc physical simulation during inference to perform fallback sampling, which is computationally inefficient. To overcome this limitation, the authors propose a reinforcement learning–based generative framework that internalizes physical validity constraints directly into the training process. The method employs an assembly-level multi-objective reward mechanism that jointly optimizes collision avoidance, global connectivity, structural interlocking, and shape consistency to guide the autoregressive policy. This approach achieves, for the first time, the generation of physically stable structures without requiring fallback strategies, delivering state-of-the-art generation quality while accelerating inference by several orders of magnitude.

Abstract

While autoregressive models have advanced 3D generation, creating physically stable brick structures remains a challenge due to the strict requirements of gravity and interconnectivity. Existing approaches rely on external physical simulators during inference to perform rejection sampling and brick-by-brick rollbacks, which severely bottlenecks efficiency. To address this, we propose a reinforcement learning paradigm that shifts physical validity enforcement from test-time correction to training-time policy optimization. By utilizing assembly-level rewards, the model optimizes for collision avoidance, global connectivity, structural interlocking, and shape conformity. This paradigm allows the model to internalize physical priors, enabling the first rollback-free generation of stable brick structures. Experimental results demonstrate that our approach achieves state-of-the-art generation quality while accelerating inference speed by orders of magnitude. Our code and dataset are available at https://github.com/miniHuiHui/STABLE. Our models are available at https://huggingface.co/miniHui/STABLE.

Problem

• physically stable
• brick structures
• autoregressive generation
• rollback-free
• 3D generation

Innovation

• rollback-free generation
• reinforcement learning
• physical stability
• brick structure synthesis
• assembly-level rewards