RoboPIN: Grounded Embodied Reasoning via Pinned Chain-of-Thought

2026-06-14

AI Summary

This work addresses the limitations of existing vision-language models in embodied reasoning, which often suffer from ambiguous entity references, cross-view inconsistencies, and a disconnect between reasoning steps and final answers due to insufficient grounding in visual evidence. To mitigate these issues, we propose the Pinned Chain-of-Thought (PinCoT) paradigm, which anchors each reasoning step to structured visual cues—including entity names, unique IDs, view indices, and spatial coordinates—to enable stable entity tracking across steps and viewpoints. We introduce a novel reasoning anchoring mechanism and a process-supervision reward to construct the first high-quality automatic dataset in PinCoT format, and employ a three-stage post-training strategy that integrates embodied knowledge injection, structured reasoning training, and alignment optimization. Our 4B-parameter model outperforms Mimo-Embodied, the strongest open-source 7B model, by an average of 12% across 14 embodied reasoning benchmarks, significantly improving grounding accuracy and identity consistency.

Abstract

Embodied reasoning requires models to perceive task-relevant objects and spaces in physical environments and maintain consistent visual grounding throughout multi-step reasoning. However, current vision-language models rely on text-only or coordinate-augmented chain-of-thought, where entity references remain implicit and ambiguous. This may cause the reasoning process to decouple from visual evidence, entity references to drift across steps, and a causal disconnection between the reasoning trajectory and the final answer, with these problems further amplified in multi-view scenarios due to cross-view appearance changes. To address these issues, we propose Pinned Chain-of-Thought (\pincot{}), a structured reasoning paradigm that pins every reasoning step to visual evidence. \pincot{} introduces the concept of \reasoninganchor{}, which binds each task-relevant entity to a structured visual anchor with entity name, unique identity, view index, and spatial grounding, enabling consistent entity tracking across reasoning steps and views. We build a fully automated data generation pipeline to construct \dataset{}, a high-quality \pincot{}-formatted reasoning dataset. We then train \method{} through three-stage post-training that progressively injects embodied knowledge, structured reasoning ability, and process-supervised alignment, with rewards that directly constrain both anchor localization and identity consistency during reasoning. On 14 benchmarks covering embodied spatial reasoning, multi-view reasoning, and pointing, \method{} with only 4B parameters consistently outperforms 7B level open-source embodied models, achieving a 12\% average improvement over the strongest 7B baseline, Mimo-Embodied. Further analysis shows that \pincot{} improves grounding accuracy and cross-step identity consistency, validating the effectiveness of process supervision.

Problem

• embodied reasoning
• visual grounding
• chain-of-thought
• entity reference drift
• multi-view reasoning

Innovation

• Pinned Chain-of-Thought
• visual grounding
• embodied reasoning
• reasoning anchor
• process supervision