Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning

2025-10-31

AI Summary

Large Vision-Language Models (LVLMs) exhibit significant deficiencies in spatial understanding; existing approaches rely on costly manual annotations, specialized sensors, or constrained environments—limiting scalability. To address this, we propose Spatial-SSRL, a self-supervised reinforcement learning paradigm that requires no human annotation. Spatial-SSRL introduces five verifiable pretraining tasks—image patch reordering, flip detection, inpainting, depth ordering, and 3D position prediction—to jointly model 2D and 3D spatial structure in a unified framework. Leveraging only standard RGB or RGB-D images, it generates reliable supervisory signals without external supervision. Evaluated on seven spatial understanding benchmarks, Spatial-SSRL improves over the Qwen2.5-VL baseline by +4.63% (3B) and +3.89% (7B), substantially enhancing spatial reasoning while preserving general visual-language comprehension capabilities.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Spatial understanding remains a weakness of Large Vision-Language Models (LVLMs). Existing supervised fine-tuning (SFT) and recent reinforcement learning with verifiable rewards (RLVR) pipelines depend on costly supervision, specialized tools, or constrained environments that limit scale. We introduce Spatial-SSRL, a self-supervised RL paradigm that derives verifiable signals directly from ordinary RGB or RGB-D images. Spatial-SSRL automatically formulates five pretext tasks that capture 2D and 3D spatial structure: shuffled patch reordering, flipped patch recognition, cropped patch inpainting, regional depth ordering, and relative 3D position prediction. These tasks provide ground-truth answers that are easy to verify and require no human or LVLM annotation. Training on our tasks substantially improves spatial reasoning while preserving general visual capabilities. On seven spatial understanding benchmarks in both image and video settings, Spatial-SSRL delivers average accuracy gains of 4.63% (3B) and 3.89% (7B) over the Qwen2.5-VL baselines. Our results show that simple, intrinsic supervision enables RLVR at scale and provides a practical route to stronger spatial intelligence in LVLMs.

Problem

• Improving spatial understanding in large vision-language models
• Reducing dependency on costly supervised training methods
• Developing self-supervised reinforcement learning for spatial reasoning

Innovation

• Self-supervised reinforcement learning from RGB images
• Automated pretext tasks for spatial structure learning
• Verifiable signals without human annotation requirements