ASPEN: Adaptive Semantic Reasoning via Command Persistence Prediction for Efficient Navigation VLAs

Vieri N. Akustia; Bin Hu

cs.ROOct 26, 2026

ASPEN: Adaptive Semantic Reasoning via Command Persistence Prediction for Efficient Navigation VLAs

Authors: Vieri N. Akustia, Bin Hu

Organizations: Department of Electrical and Computer Engineering, University of Houston, Houston, TX 77004, USA · Departement of Engineering Technology, University of Houston, Houston, TX 77004, USA

Abstract

Navigation vision-language-action (VLA) policies repeatedly invoke costly semantic reasoning even when the next command remains unchanged. Command reuse can reduce computation but risks missing instruction-dependent turns or stops. We introduce ASPEN, an event-triggered scheduler based on command persistence prediction: estimating whether the frozen base policy would reissue its previous command. A lightweight multimodal predictor combines temporal observations, language instructions, execution history, and cached VLA confidence. At eligible command boundaries, sufficient predicted persistence permits command reuse. Otherwise, ASPEN triggers fresh semantic reasoning from the VLA. Evaluated with frozen NaVILA on cross-dataset RxR-CE val-unseen, ASPEN reduces VLA calls by 9.5% and estimated inference time by 8.1%, including scheduling overhead. It maintains comparable navigation performance and preserves identical baseline action sequences on 92.1% of episodes.

Figures & tables

Figure 1: ASPEN schedules a frozen navigation VLA. A learned persistence score selects command reuse or fresh VLA inference, preceded by an eligibility check.

Method	VLA calls/ep (vs baseline)	Inference s/ep (vs baseline)	SR $\uparrow$	OS $\uparrow$	nDTW $\uparrow$	NE $\downarrow$	Exact agreement (%) $\uparrow$
Baseline NaVILA	75.9	41.5	53.0	59.9	63.3	5.93	100.0%
FlashTrigger Tan et al. (2025)	72.0 ( $-5.1\%$ )	39.7 ( $-4.5\%$ )	53.0	59.9	63.3	5.93	97.5%
SP-VLA Li et al. (2026)	69.4 ( $-8.5\%$ )	38.0 ( $-8.5\%$ )	42.7	53.2	53.2	7.17	7.3%
Free Rule (ours)	61.0 ( $-19.6\%$ )	33.4 ( $-19.6\%$ )	49.1	57.3	58.4	6.64	7.1%
Full Gate (ours)	66.9 ( $-11.7\%$ )	40.2 ( $-3.3\%$ )	51.6	59.5	61.8	6.36	21.8%
ASPEN (ours)	68.7 ( $-9.5\%$ )	38.2 ( $-8.1\%$ )	53.1	59.8	63.2	5.95	92.1%

Table 1: Cross-dataset evaluation on 1,223 RxR-CE val-unseen episodes. Parentheses show changes relative to baseline NaVILA.

Figure 2: Example RxR-CE val-unseen episode with baseline action sequence preservation. ASPEN reduces NaVILA calls from 53 to 47 (11%) and estimated inference time from 29.0 to 26.4 s (9%).

Figure 3: A critical command boundary in an example RxR-CE val-unseen episode. ASPEN refreshes semantic reasoning and turns right, whereas the navigation-adapted SP-VLA continues forward and subsequently fails.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Efficiency–performance trade-offs on RxR-CE val-unseen. Each point represents a method in Table 1 . Both panel compare estimated inference time per episode with (a) success rate and (b) identical trajectory rate.

Efficient VLA Model Inference

Efficient vision-language-action (VLA) model inference reduces robot-control latency, memory use, and energy while preserving task performance.

See all 190 papers

NavJev: Efficient Vision-Language Navigation via Action-Centric Visual Compression and Discriminative Action-Semantic Memory
Sep 28, 2026Kai Sheng, Liuyi Wang, Jinlong Li +3Efficient VLM Inference Vision-Language Navigation
College of Electronic and Information Engineering, Tongji University, Shanghai, China
SpaAct: Spatially-Activated Transition Learning with Curriculum Adaptation for Vision-Language Navigation
Apr 30, 2026Pengna Li, Kangyi Wu, Shaoqing Xu +7VLM Adaptation Vision-Language Navigation
National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University · The State Key Laboratory of Internet of Things for Smart City, Centre for Artificial Intelligence and Robotics, University of Macau · Xiaomi EV +2
VerNav: Verifier-First Low-Latency Vision-and-Language Navigation
Sep 1, 2026Zhixin Wang, Chengzheyi Yao, Leyuan Liu +2Efficient VLM Inference Vision-Language Navigation
University of Electronic Science and Technology of China

Sep 28, 2026J/K move · Enter open · S save

NavJev: Efficient Vision-Language Navigation via Action-Centric Visual Compression and Discriminative Action-Semantic Memory

Kai Sheng, Liuyi Wang, Jinlong Li, Haojie Dai+2

College of Electronic and Information Engineering, Tongji University, Shanghai, China

Efficient VLM Inference Vision-Language Navigation

Open paper AskPDF arXiv

Recent zero-shot Vision-and-Language Navigation (VLN) methods increasingly rely on multimodal large language models (MLLMs) to reason over visual observations, navigation instructions, and candidate actions. Although effective, repeatedly invoking autoregressive multimodal reasoning at every navigation step introduces substantial inference latency, limiting the responsiveness of embodied agents. We propose NavJev, an efficient VLN framework that reformulates online navigation from repeated multimodal generation into compact visual compression followed by lightweight typed action selection. Specifically, Action-Centric Visual Compression (ACVC) integrates waypoint geometry, BLIP captions, and RAM semantic tags into compact representations of candidate actions, while Discriminative Action-Semantic Memory (DASM) filters shared semantics and maintains discriminative action-specific evidence across navigation steps. Based on these representations, Jev directly performs structured probabilistic decisions over the available action set. Experiments on R2R-CE show that NavJev achieves 27.0% SR and 22.4% SPL with only 0.65 s per navigation step, while substantially reducing inference latency and cost compared with MLLM-based VLN methods. The project page is available at https://kai-sheng-caesar.github.io/NavJev/.

Abstract

Efficient VLA Model Inference

Explore similar work

NavJev: Efficient Vision-Language Navigation via Action-Centric Visual Compression and Discriminative Action-Semantic Memory

SpaAct: Spatially-Activated Transition Learning with Curriculum Adaptation for Vision-Language Navigation

VerNav: Verifier-First Low-Latency Vision-and-Language Navigation