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

TL;DR

The paper introduces Spatial-SSRL, a self supervised reinforcement learning framework that enhances spatial understanding in LVLMs by deriving verifiable supervision from raw RGB or RGB-D images. It defines five pretext tasks split into depth free and depth based categories, and trains models via a two stage process with GRPO using verifiable rewards. A self supervised Spatial-SSRL-81k dataset enables zero label generation and scalable RL optimization, achieving notable gains on seven spatial benchmarks while preserving general visual capabilities. The approach demonstrates strong 3D reasoning improvements, cross modal transfer to video, and practical scalability due to its tool free design and intrinsic supervision. This work provides practical evidence that self supervised signals can meaningfully boost spatial intelligence in LVLMs with broad implications for robotics, autonomous systems, and embodied AI.

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.

Figures (11)

• Figure 1: We present Spatial-SSRL, a self-supervised reinforcement learning paradigm for spatial understanding. (a) Qualitative examples: the baseline answers are wrong (red), whereas our model predicts correctly (green) for 3D locations and orientations. (b) Quantitative results on seven spatial benchmarks show consistent improvements of Spatial-SSRL-7B against Qwen2.5-VL-7B and its CoT variant.
• Figure 2: (a) Prior pipelines boost spatial understanding by injecting extrinsic supervision from expert tools or synthetic environments, which inflates cost and limits scalability. (b) Our Spatial-SSRL replaces these dependencies with intrinsic self-supervision, yielding a scalable, lightweight, low-cost, and naturally verifiable pipeline.
• Figure 3: Overview of Spatial-SSRL.(a) Self-supervised data curation: from raw RGB and RGB-D images, we automatically construct five pretext tasks (patch reordering, patch flip detection, cropped-patch inpainting, regional depth ordering, and relative 3D position prediction), requiring no human or LLM annotations. (b) RL training: the model is optimized with Group Relative Policy Optimization (GRPO) using a verifiable reward function that evaluates answer correctness, and a format reward that elicits format compliance.
• Figure 4: Examples of the task Shuffled Patch Reordering.
• Figure 5: Examples of the task Flipped Patch Recognition.