HiRO-Nav: Hybrid ReasOning Enables Efficient Embodied Navigation

He Zhao; Yijun Yang; Zichuan Lin; Deheng Ye; Chunyan Miao

cs.AIApr 9, 2026

HiRO-Nav: Hybrid ReasOning Enables Efficient Embodied Navigation

Authors: He Zhao, Yijun Yang, Zichuan Lin, Deheng Ye, Chunyan Miao

Organizations: Nanyang Technological University · Tencent · Independent Researcher

Abstract

Embodied navigation agents built upon large reasoning models (LRMs) can handle complex, multimodal environmental input and perform grounded reasoning per step to improve sequential decision-making for long-horizon tasks. However, a critical question remains: how can the reasoning capabilities of LRMs be harnessed intelligently and efficiently for long-horizon navigation tasks? In simple scenes, agents are expected to act reflexively, while in complex ones they should engage in deliberate reasoning before acting.To achieve this, we introduce Hybrid ReasOning Navigation (HiRO-Nav) agent, the first kind of agent capable of adaptively determining whether to perform thinking at every step based on its own action entropy. Specifically, by examining how the agent's action entropy evolves over the navigation trajectories, we observed that only a small fraction of actions exhibit high entropy, and these actions often steer the agent toward novel scenes or critical objects. Furthermore, studying the relationship between action entropy and task completion (i.e., Q-value) reveals that improving high-entropy actions contributes more positively to task success.Hence, we propose a tailored training pipeline comprising hybrid supervised fine-tuning as a cold start, followed by online reinforcement learning with the proposed hybrid reasoning strategy to explicitly activate reasoning only for high-entropy actions, significantly reducing computational overhead while improving decision quality. Extensive experiments on the CHORES- $\mathbb{S}$ ObjectNav benchmark showcases that HiRO-Nav achieves a better trade-off between success rates and token efficiency than both dense-thinking and no-thinking baselines.

Figures & tables

Figure 1: Illustration of HiRO-Nav agent adaptively determining whether to perform thinking based on its own action entropy $H_{\pi_{\theta}}(a_{t},\bm{x}_{t})$ . Based on our observations, high-entropy actions often steer the agent toward novel scenes or critical objects, which are located on key waypoints over the navigation trajectory, i.e., the black dashed line in the bird-eye-view map. HiRO-Nav accordingly activates reasoning only for these high-entropy actions (red points), improving the trade-off between reasoning efficiency and performance compared to dense-thinking and no-thinking agents, as demonstrated in Fig. 4 .

Panel of Figure 2 — Figure 2: The distribution of action entropy (AE) over navigation trajectories . We analyze the AE distribution of a VLM agent fine-tuned using expert trajectories on CHORES - $\mathbb{S}$ ObjectNav tasks. (a): Only a small fraction ( $\sim$ 30%) of actions exhibits high entropy (AE $\geq$ 0.6). (b): High-entropy actions (red points in the map) often steer the agent to explore novel areas or approach critical objects. An extended version of the figure (b) can be found in Fig. 10 of Appendix.

Figure 3: Mean Q-value of a hybrid fine-tuned model introduced in Sec. 3.3 across various action entropy thresholds . The high threshold means sparse activation of reasoning, resulting in high token efficiency. We conclude that thinking only for high-entropy actions (threshold

=

0.6) achieves the best trade-off between task completion and maximizing token efficiency. Lower or higher thresholds can result in “overthinking” or “underthinking” respectively, both degrading the final performance.

Figure 4: Comparison of HiRO-Nav agent against SOTA baselines in terms of the trade-off between navigation success rate (SR) and token efficiency . We compute the average number of model-generated tokens per episode (#Token/E). HiRO-Nav with hybrid reasoning strategy (Ours) achieves the best trade-off.

Figure 5: Overview of HiRO-Nav training pipeline and the proposed hybrid reasoning strategy. which consists of two part: (1) Hybrid supervised fine-tuning (left). We first collect a hybrid reasoning dataset (HRD) containing a no-reasoning dataset and a reasoning dataset annotated by Gemini2.0-Flash on high-entropy actions . We then fine-tune a VLM on the HRD as a cold start to enable the agent with hybrid-thinking abilities. (2) Two-stage online reinforcement learning (bottom right). In stage I, we collect no-thinking rollouts and train the agent’s no-thinking ability. In Stage II, we collect rollouts with proposed hybrid reasoning strategy and train agent’s thinking ability. The proposed Hybrid reasoning strategy (top right) encourages the agent to activate thinking only for high-entropy actions.

Training Recipe	Model Name	Reasoning Strategy	ASM	Dataset	SR% $\uparrow$	SEL% $\uparrow$	#Token/E $\downarrow$	#Token/S $\downarrow$
TF	GPT-4o [ 18 ]	No-Thinking	$\checkmark$	-	8.0	4.4	$5.6\times 10^{2}$	3.1
	o3 [ 22 ]	Dense-Thinking	$\checkmark$	-	27.0	12.3	$2.4\times 10^{5}$	539.4
	Gemini2.5-Pro [ 9 ]	Dense-Thinking	$\checkmark$	-	27.5	16.2	$2.3\times 10^{5}$	804.8
SFT	NaVILA [ 7 ]	No-Thinking	$\times$	NRD	30.5	12.1	$1.8\times 10^{3}$	3.7
	Qwen2.5VL-3B [ 1 ]	No-Thinking	$\times$	NRD	36.5	28.8	$1.5\times 10^{3}$	3.8
	Qwen2.5VL-3B [ 1 ]	No-Thinking	$\checkmark$	NRD	50.0	41.2	$1.2\times 10^{3}$	3.8

Table 1: Comparison with state of the art. SR = success rate, SEL = success rate weighted by episode length, Ours = our proposed hybrid reasoning strategy, NRD = no-reasoning dataset, HRD = hybrid-reasoning dataset, TF = training-free.

K

in the Thinking-Every-

K

-Steps baseline and the size of no-thinking window (NTW) are equally set to 5 for a fair comparison.

\checkmark

\times

refers to the corresponding annotated semantic map (ASM) included/not included in the input of agents. #Token/E and #Token/S denote the number of tokens generated by the agent per episode and per step, respectively. RL

{}_{\text{I}}

and RL

{}_{\text{II}}

refer to the two stages of online RL training.

Figure 6: Comparison of reasoning efficiency of hybrid reasoning (Ours) and thinking-every-

\bm{K}

-steps (Every-

\bm{K}

) . We divide the navigation tasks into different difficulty levels based on the ground truth shortest path lengths. Our hybrid reasoning method consistently outperforms the baseline reasoning approach across all difficulty levels while maintaining a lower thinking ratio. TR=Thinking Ratio. SR= Success Rate.

Panel of Figure 7 — Figure 7: Ablation study. Fig. (a)&(b) : Vanilla RL fails to effectively enhance hybrid reasoning ability due to a decline in no-thinking ability. In contrast, our two-stage training paradigm successfully improves the agent’s no-thinking ability in Stage I and maintained by KL constraint in Stage II, which subsequently enhance the hybrid reasoning ability. Fig. (c) : The superior performance of NTW>0 highlights its effectiveness. Our hybrid reasoning enables the agent to achieve improved performance with a reduced thinking ratio (indicated by high $\tau$ ), reaching a SOTA result at $\tau$ =0.6 and NTW=5.

ASM	Reasoning Strategy	SR%	SEL%
GT	Dense-Thinking	34.5	15.6
	No-Thinking	70.0	48.8
	Ours	81.0	57.2
DM	Ours	72.0	46.8

Table 2: Robustness analysis of HiRO-Nav using annotated semantic maps constructed by deep models (DM) in place of ground truth (GT) ASMs. We apply Depth-Anything [ 38 ] to estimate depth information and a fine-tuned Mask R-CNN [ 16 ] to perform instance segmentation during ASM construction. Despite of performance drop due to noise introduced by inaccurate estimation, HiRO-Nav with hybrid reasoning still outperforms no-thinking with ground truth ASMs, showing its robustness in resisting noise in ASM.

Figure 8: Pass@k curves of HiRO-Nav with hybrid reasoning. GT and DM refer to the ground truth ASMs and the ASMs estimated by deep models as in Tab. 2 .We evaluation for 16 times with temperature=0.2. The navigation ability upper bound of HiRO-Nav outperforms task-specific sota method Poliformer [ 43 ] , even when using noisy deep model estimated ASMs.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: An example of annotated semantic map.

Action	Argument
move_ahead	0.2 meter
move_back	0.2 meter
rotate_left	30 degree
rotate_right	30 degree
rotate_left_small	5 degree
rotate_right_small	5 degree

Appendix

Table 3: Action Space and corresponding arguments. All actions are in textual form.

Figure 10: Additional visualization examples of action entropy at each navigation waypoint.

Panel of Figure 11 — Figure 11: Two action entropy calculation method show the same relationship between action entropy and Q-value.

Model	Action Entropy	SR%	SEL%
Qwen2.5VL-3B ${}_{\text{H-SFT}}$	First Token	59.5	48.9
Qwen2.5VL-3B ${}_{\text{H-SFT}}$	Mean	58.0	43.1
HiRO-Nav	First Token	81.0	57.2
HiRO-Nav	Mean	79.5	53.1

Appendix

Table 4: We compare using first-token entropy versus mean token entropy as measures of action entropy. The performance difference between the two methods is marginal.

Adaptive Reasoning in Language Models

See all 13 papers

Embodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation
Aug 18, 2026Hongyan Feng, Sunlai Chen, Xuanyu Liu +9Embodied Navigation Vision-Language Navigation
School of Software Technolog,Zhejiang University · Zhejiang Humanoid Robot Innovation Center Co., Ltd.
Lifelong Embodied Navigation Learning
Mar 6, 2026Xudong Wang, Jiahua Dong, Baichen Liu +3Embodied Navigation Continual Learning for LLM Agents
State Key Laboratory of Robotics and Intelligent Systems, Shenyang Institute of Automation, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Mohamed bin Zayed University of Artificial Intelligence
PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation
Oct 4, 2026Jing Xie, Shouwei Ruan, Yubin Wang +4Embodied Navigation Robot Navigation
MoE Key Lab of Artificial Intelligence, Institute of AI, School of Computer Science, Shanghai Jiao Tong University, Shanghai, China · Beihang University, Beijing, China · The Hong Kong University of Science and Technology, Hong Kong, China +3

Aug 18, 2026J/K move · Enter open · S save

Embodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation

Hongyan Feng, Sunlai Chen, Xuanyu Liu, Miao Pan+8

School of Software Technolog,Zhejiang University · Zhejiang Humanoid Robot Innovation Center Co., Ltd.

Embodied Navigation Vision-Language Navigation Visual Navigation Memory-Augmented VLMs RL for VLMs

Open paper AskPDF arXiv

Although Large Vision-Language Models (VLMs) have significantly advanced embodied navigation, their direct deployment remains challenging, as existing methods often force VLMs into unnatural action spaces that misalign with their 2D pre-training priors, compounded by rigid reasoning schedules and inefficient memory management. To overcome these limitations, we propose TAMP-Nav, a unified framework for efficient embodied navigation. First, we introduce a Pixel-to-3D Action Formulation (Point) that reformulates navigation into 2D visual prompting. Specifically, the VLM merely selects 2D pixels, which are then projected into 3D coordinates for a low-level SLAM controller. This design naturally aligns embodied execution with the VLM's inherent 2D visual capabilities. Second, we propose an integrated Selective Reasoning and Anchor-Trajectory Memory mechanism (Think and Memorize), which dynamically triggers Chain-of-Thought and retains high-fidelity memory only at critical nodes, compressing redundant trajectories into lightweight Space-Time Indicators, thereby preserving critical historical information and enhancing spatio-temporal perception. Finally, we design an efficient Two-Level Alignment Paradigm (Align) via Group Relative Policy Optimization (GRPO). By superimposing global outcome rewards with fine-grained process rewards, this dense supervision tightly aligns the agent's cognitive planning with physical environmental feedback, endowing the model with adaptive reasoning capabilities. Experiments demonstrate that TAMP-Nav achieves state-of-the-art performance (e.g., 66.2% SR on R2R-CE) with high runtime and sample efficiency (requiring only 90k training trajectories).

Abstract

Adaptive Reasoning in Language Models

Explore similar work

Embodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation

Lifelong Embodied Navigation Learning

PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation