Technical Report of RoboSpatial Challenge at CVPR 2026: Selective Reasoning Activation and Reference-Frame Disambiguation for Embodied Spatial Reasoning

2026-06-30

AI Summary

Existing vision-language models often struggle with spatial understanding in embodied tasks due to ambiguities in reference frames and insufficient reasoning capabilities. This work proposes RoboSpatialBrain, a system built upon RoboBrain2.5-8B-NV that enhances spatial reasoning without any model training. It achieves this by prompting deliberate reasoning through a forced <think> prefix combined with task-specific post-prompts, and by introducing an explicit reference frame realignment mechanism that disambiguates between camera-centric and object-centric perspectives. Additionally, the study explores a compatibility-aware data fine-tuning strategy. Without altering the model architecture, the approach substantially improves performance, achieving an 80.9% success rate on the RoboSpatial-Home benchmark and winning the CVPR 2026 RoboSpatial Challenge. This work also provides the first systematic analysis of the interaction between prompt engineering and fine-tuning in spatial reasoning tasks.

Abstract

Vision-language models achieve strong general perception but often struggle with the spatial reasoning required for embodied tasks. We present RoboSpatialBrain, our submission to the RoboSpatial Challenge at the Embodied Reasoning in Action Workshop, CVPR 2026, built on RoboBrain2.5-8B-NV. RoboSpatialBrain combines two training-free, inference-time mechanisms: a forced <think> prefix activation strategy paired with a task-specific post-prompt that elicits deliberate reasoning on context and compatibility tasks, and an explicit reference-frame redirection pipeline that resolves camera-centric and object-centric ambiguity for context tasks. We additionally explore fine-tuning RoboBrain2.5 on compatibility data and present a detailed analysis of its interaction with prompting. RoboSpatialBrain achieved first place in the RoboSpatial Challenge, with an overall success rate of 80.9\% on RoboSpatial-Home. Code is available at https://github.com/YuxiangXie2003/RoboSpatialBrain.

Problem

• spatial reasoning
• embodied tasks
• reference-frame ambiguity
• vision-language models

Innovation

• Selective Reasoning Activation
• Reference-Frame Disambiguation
• Embodied Spatial Reasoning
• Prompt Engineering
• Vision-Language Models