OctoNav：迈向通用具身导航
OctoNav: Towards Generalist Embodied Navigation
把 MLLM 改造成仅凭 2D 视觉观测输出低层动作的 VLA 模型，并用 TBA-CoT 数据加 GRPO 做「先思考后行动」的强化学习训练，为视觉研究者提供了将推理链注入具身导航策略的可复用范式。
原文依据 adapt it to a VLA-type model, which can produce low-level actions solely based on 2D visual observations
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论具身导航是追求具身智能的基础支柱之一。然而，以往的导航研究被划分为不同的任务/能力，例如 ObjNav、ImgNav 和 VLN，它们在任务设置/目标和模态上各不相同，导致数据集与方法各自独立设计。在本工作中，我们朝着通用导航迈出步伐，该导航能够遵循包含任意模态与能力组合的自由形式指令。为实现这一目标，我们提出了一个大规模基准及相应方法，分别称为 OctoNav-Bench 和 OctoNav-R1。具体而言，OctoNav-Bench 通过一个设计好的自动标注流程构建。我们精心制作了指令-轨迹对，其中指令在自由形式上具有多样性，涵盖任意模态与能力。此外，我们在 OctoNav-Bench 内构建了一个先思考后行动（TBA-CoT）数据集，以提供动作背后的思考过程。对于 OctoNav-R1，我们将其建立在 MLLM 之上，并将其改造为 VLA 型模型，该模型能够仅基于 2D 视觉观测产生低层动作。此外，我们设计了一种混合训练范式（HTP），它包含三个阶段，即 Action-/TBA-SFT、Nav-GRPO 和在线 RL 阶段。每个阶段都包含设计好的学习策略与奖励。具体而言，受 OpenAI-o1 和 DeepSeek-R1 的启发——它们通过先思考后回答展现出令人印象深刻的推理能力——我们设计了 TBA-SFT 和 Nav-GRPO，以在具身导航中实现先思考后行动，从而提升模型面向通用智能体的推理能力。TBA-SFT 利用 TBA-CoT 数据集对模型进行微调，随后我们借助 Nav-GRPO 提升其思考能力。最终，OctoNav-R1 相比以往方法展现出更优的性能。
英文原文摘要 1772 字符
Embodied navigation stands as a foundation pillar within the pursuit of embodied intelligence. However, previous navigation research is divided into different tasks/capabilities, e.g., ObjNav, ImgNav and VLN, where they differ in task settings/objectives and modalities, making datasets and methods designed individually. In this work, we take steps toward generalist navigation, which can follow free-form instructions that include arbitrary compounds of modality and capability. To achieve this, we propose a large-scale benchmark and corresponding method, termed OctoNav-Bench and OctoNav-R1. Specifically, OctoNav-Bench is constructed via a designed automatic annotation pipeline. We thoroughly craft instruction-trajectory pairs, where instructions are diverse in free-form with arbitrary modality and capability. Also, we construct a Think-Before-Action (TBA-CoT) dataset within OctoNav-Bench to provide the thinking process behind actions. For OctoNav-R1, we build it upon MLLMs and adapt it to a VLA-type model, which can produce low-level actions solely based on 2D visual observations. Moreover, we design a Hybrid Training Paradigm (HTP) that consists of three stages, i.e., Action-/TBA-SFT, Nav-GRPO, and Online RL stages. Each stage contains designed learning policies and rewards. Specifically, inspired by the OpenAI-o1 and DeepSeek-R1, which show impressive reasoning ability via thinking-before-answer, we design TBA-SFT and Nav-GRPO to achieve thinking-before-action for embodied navigation, improving model's reasoning ability toward generalists. TBA-SFT utilizes the TBA-CoT dataset to fine-tune the model, and then we leverage Nav-GRPO to improve its thinking ability. Finally, OctoNav-R1 shows superior performance compared with the previous methods.