Skip to content

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

🕒 Published (v1): 2026-06-30 13:25 UTC · Source: Arxiv · link

Why this paper was selected

CVPR 2026; embodied spatial reasoning benchmark + winning solution

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

RoboSpatialBrain is the first-place solution (80.9% overall success rate) to the RoboSpatial Challenge at CVPR 2026, built on RoboBrain2.5-8B-NV. It uses two training-free inference-time mechanisms—selective <think> prefix activation and a reference-frame redirection pipeline—to improve embodied spatial reasoning without any model fine-tuning in the final pipeline.

Problem

VLMs achieve strong general perception but consistently fail at embodied spatial reasoning tasks, particularly: (1) reasoning about vacant space and metric compatibility for object placement, and (2) reference-frame ambiguity, where models default to camera-centric interpretations of spatial language when queries require an object-centric frame of reference.

Method

The pipeline routes each query by task type using a lightweight Qwen3.5-2B classifier, then applies two mechanisms:

Selective reasoning activation: For reasoning-required tasks (context and compatibility), a <think> prefix is forced onto the model's output token stream and a task-specific post-prompt is injected. Since RoboBrain2.5 (built on Qwen3-VL) has no explicit thinking-mode toggle, forcing the prefix is hypothesized to implicitly activate residual reasoning-format priors from the Qwen pretraining lineage. Configuration tasks (reasoning-free) receive no intervention to avoid overthinking degradation.

Reference-frame redirection: For context tasks, Qwen3.5-2B extracts the reference object from the query. RoboBrain2.5 is then prompted to determine that object's facing direction. A precomputed direction mapping table rewrites the object-centric spatial query into its camera-centric equivalent, and RoboBrain2.5 is queried again on the rewritten version. Final predictions are obtained by merging answers from both the original and rewritten queries. No fine-tuning is required.

Fine-tuning on 24,000 compatibility-task VQA examples constructed from EmbodiedScan was explored (LoRA rank 16, 4×A100, AdamW, \(\text{lr}=1\times10^{-4}\)) but excluded from the final pipeline due to interference with prompt-based reasoning elicitation.

Key Contributions

  • Selective <think> prefix activation that routes reasoning elicitation by task type, yielding a +43.8 pp gain on compatibility tasks
  • Training-free reference-frame redirection pipeline that consistently resolves camera-centric vs. object-centric ambiguity across all prompting configurations
  • Empirical analysis showing that fine-tuning with narrow-domain data degrades the model's general reasoning ability, undermining prompt-based elicitation—a manifestation of catastrophic forgetting
  • First-place submission (80.9% overall) to the RoboSpatial Challenge at CVPR 2026 ERA Workshop

Results

  • Overall success rate: 80.9% on RoboSpatial-Home (first place)
  • Configuration task: 88.6% (best combination: no post-prompt, no <think>)
  • Compatibility task: 83.8% (best: post-prompt + <think>; base without either: 40.0%, a +43.8 pp gain)
  • Context task: 70.5% (best: post-prompt + <think> + reference-frame redirection)
  • Reference-frame redirection adds +1.7 to +3.3 pp on context tasks across all prompting configurations
  • Fine-tuning at 1,000 steps without prompting: +27.6 pp over base; but with post-prompt + <think>: −12.4 pp vs. base, worsening to −22.8 pp at 3,000 steps

Limitations

  • Reference-frame redirection relies on a precomputed direction mapping table; coverage of unusual or non-canonical objects is unclear
  • Fine-tuning experiments used a single data source (EmbodiedScan) with limited scene and question diversity, making findings on SFT-prompt interaction potentially dataset-specific
  • Context task performance (70.5%) lags compatibility (83.8%) and configuration (88.6%), indicating residual difficulty in vacant-space localization
  • The mechanism by which forced <think> prefix activates internal reasoning pathways is hypothesized but not empirically verified at the representation level
  • Final pipeline is entirely inference-time; no improvement to the backbone weights means performance is bounded by RoboBrain2.5's pretrained spatial priors

Relevance to Vision-Language Models

This work directly addresses a known weakness of VLMs—spatial reasoning and reference-frame grounding—using inference-time prompt engineering rather than retraining, making it broadly applicable to any VLM with residual chain-of-thought priors. The selective activation finding (thinking helps complex tasks, hurts simple ones) is an actionable insight for anyone deploying reasoning-capable VLMs on heterogeneous task distributions. The reference-frame redirection pipeline exemplifies how LLM-mediated query rewriting can compensate for systematic VLM biases without modifying model weights. The catastrophic forgetting interaction with prompt elicitation is a cautionary result for VLM fine-tuning pipelines that also rely on prompt-based reasoning augmentation.