Cognitive-Horizon Action Chunking: Memory-Guided Adaptive Reasoning for Embodied Navigation
Authors: Weile Gong, Zijian Lu, Jiahao Zheng, Bo Xu, Yiping Zuo, Jiajia Guo, Xin He, Weijun Wang, Weibei Fan
Organizations: Nanjing University of Posts and Telecommunications · National Mobile Communications Research Laboratory, Southeast University · Institute for AI Industry Research, Tsinghua University
Multimodal foundation models are advancing embodied navigation, but frequent inference incurs substantial computational cost. We present Cognitive-Horizon Action Chunking (CHAC), a training-free framework for memory-guided adaptive reasoning with a frozen navigation policy. Retained visual context, completion cues, and controller feedback determine the cognitive horizon, or how many cached waypoints the robot executes before refreshing its prediction. At the policy's stop request, local visitation history guides a bounded correction with at most one extra movement and no additional model calls. CHAC reduces vision-language model calls by 33.4% on HM3D-OVON and 64.7% on R2R-CE, with success-rate decreases of 0.63 and 10.82 percentage points, respectively. Success rate and success weighted by path length improve per model call on both benchmarks. Across four physical navigation tasks with 40 runs per method, CHAC achieves 18 successes versus 14 for native execution, reducing model calls by 54.9% and GPU inference time by 49.7%.
Figures & tables
Figure 1: Overview of CHAC. A frozen navigation policy predicts ordered waypoints from current views, visual history, and the task instruction. During execution, retained visual context and commitment checks guide when to refresh the cached prediction. At native STOP, nearby visited positions guide a bounded terminal correction without another model call. The lower scene shows how executed motion becomes visitation history. Red dots mark visited positions, and the blue disk denotes the local-history neighborhood.
Method
SR (%) ↑
SPL (%) ↑
VLM Calls ↓
FlowEq ↓
Env. Actions
ESR↑
ESPL↑
HM3D-OVON Val-Unseen [ 30 ]
OmniNav-Fast [ 28 ] (ICLR 2026)
32.17
21.92
130,444
1,304,440
130,444
1.000
1.000
AAC [ 19 ] (CVPR 2026)
14.20
5.28
84,869
16,973,800
276,726
0.679
0.376
SGAC [ 23 ] (NeurIPS 2025)
26.23
16.18
330,314
≥ 3,303,140
330,314
0.322
0.300
ChunkTrust-AHS [ 13 ] (arXiv 2026)
17.13
7.72
98,596
985,960
207,696
0.705
0.476
DVAC [ 11 ] (arXiv 2026)
7.13
1.71
91,006
910,060
350,266
0.318
0.112
Table 1: Inference demand and navigation quality. Costs are totals over each benchmark. Gray rows identify CHAC. Bold, underlined, and italic entries identify the top three SR/SPL values per benchmark. For ESR and ESPL , bold and underline identify the top two. SPL and ESPL use the common valid subset. ESR uses all episodes.
Configuration
SR (%) ↑
SPL (%) ↑
VLM Calls ↓
FlowEq ↓
Env. Actions
ESR↑
ESPL↑
OmniNav-Fast (Native)
30.60
20.47
45,286
452,860
45,286
1.000
1.000
Fixed-3
25.20
15.10
39,683
396,830
96,883
0.940
0.842
Fixed-5
21.40
11.60
33,814
338,140
116,014
0.937
0.766
Fixed-5 + Visual
21.00
11.46
35,111
351,110
118,311
0.885
0.728
Fixed-5 + Commitment
23.10
13.13
30,863
308,630
96,878
1.108
0.949
Fixed-5 + Terminal
29.40
14.77
33,814
338,140
116,943
1.287
0.975
Table 2: Component contributions on 1,000 matched HM3D episodes. Visual, Commitment, and Terminal denote the two refresh boundaries and terminal correction. Single-component rows extend Fixed-5. Removal rows start from Full CHAC. SPL uses 991 common valid episodes.
Figure 2: Global cost changes and boundary interaction. (a) Full CHAC versus w/o Commitment on all 1,000 pairs. Both retain Visual and Terminal. Peach indicates lower cost, rose indicates higher cost, coral indicates mixed changes, and gray indicates unchanged costs. (b) Adding Visual with Commitment off or on, and the difference between these effects. All four configurations disable Terminal. Error bars are paired-bootstrap 95% intervals.
Figure 3: Small displacement, changed visual evidence (episode 1644). Blue solid: CHAC; orange dashed: w/o Visual. Both retain Commitment and Terminal. Left: the anchor and refresh states are only 0.12 m apart but differ in heading by 129∘ ; purple and red mark the corresponding views. CHAC refreshes before waypoint 5, while w/o Visual executes the cached waypoint. Right: full trajectories toward the exercise bike, with dots marking VLM calls. CHAC uses 41 calls versus 47 and travels 20.64 m less.
Strategy
Rescue ↑
Regression ↓
SR (%) ↑
Original STOP
0
0
23.00
Motion continuation
12
80
16.20
Motion reversal
59
25
26.40
Local consensus
82
16
29.60
Table 3: Terminal direction under matched requested radii. All strategies share neural prefixes, native motor/STOP conditions, and zero additional VLM or Flow calls. SR includes all 1,000 episodes.
Method
Success (%)
Calls
GPU (s)
OmniNav-Fast
35.0
687
437.76
Fixed-3
40.0
357
262.54
CHAC (Ours)
45.0
310
220.00
Table 4: Physical navigation with a frozen policy. Each method runs ten times on each of four indoor tasks. Success is reported as a percentage. Calls and GPU times are totals. GPU time measures model inference.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Task
Method
Success
VLM Calls
GPU (s)
Actions
Reused
A1
OmniNav-Fast
0/10
17.7
11.57
16.9
0.0
Fixed-3
3/10
6.4
4.90
16.5
11.0
CHAC
4/10
5.6
4.16
14.5
8.1
A2
OmniNav-Fast
7/10
15.7
10.17
14.7
0.0
Fixed-3
7/10
8.8
6.08
23.4
15.6
CHAC
6/10
8.0
5.52
22.6
14.2
Appendix
Table 5: Robot navigation detail per task and method.
Figure 5: Success rate and VLM calls. All methods and values come from Table 1 . Blue stars mark CHAC; orange points mark native execution. Moving toward the upper left means higher success with fewer VLM calls. Each point represents one evaluated method configuration.
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).
Hongyan Feng, Sunlai Chen, Xuanyu Liu +9
School of Software Technolog,Zhejiang University · Zhejiang Humanoid Robot Innovation Center Co., Ltd.
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-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.
He Zhao, Yijun Yang, Zichuan Lin +2
Nanyang Technological University · Tencent · Independent Researcher
Vision-Language Navigation (VLN) requires embodied agents to generate actions based on instructions and observations. General-purpose multimodal agents offer a promising basis for this task, but selecting plausible local actions does not ensure that execution remains consistent with the intended route, particularly in long-horizon tasks. Moreover, the accumulated interaction history increases the input required for subsequent decisions, resulting in a significant inference overhead. To this end, we introduce \method, an Agentic VLN framework that includes a Goal Agent that sets adaptive goals for local actions, a Verify Agent that dynamically verifies whether a goal has been completed, a Memory Agent for multimodal context compression, and a Visuomotor Agent to execute adaptive goals. Specifically, the Goal Agent formulates adaptive goals based on the instruction, current observation, and execution history. Then the Visuomotor Agent executes navigation actions to achieve each goal, while the Verify Agent uses a goal-specific verification question to dynamically assess whether the observed outcomes satisfy the intended completion condition. Verified goal completion then marks a boundary for the Memory Agent to compress the corresponding multimodal interaction history while preserving information needed for subsequent navigation. We evaluate navigation on R2R-CE and RxR-CE, examine framework variants across three model backbones, and study context evolution during execution. For Real-World evaluation, \method achieves 83.3% success and 1.51,m navigation error across eight challenging routes evaluated three times each.
Haoxiang Shi, Zaijing Li, Muhe Ding +3
Harbin Institute of Technology (Shenzhen) · Pengcheng Laboratory