TACET: Context-Appropriate Acoustic-Social Navigation for Quadrupeds

Sungsan Park; Young-Sik Shin; Sanghyun Kim

cs.ROOct 2, 2026

TACET: Context-Appropriate Acoustic-Social Navigation for Quadrupeds

Authors: Sungsan Park, Young-Sik Shin, Sanghyun Kim

Organizations: Department of Mechanical Engineering, Kyung Hee University, Yongin, 17104, Republic of Korea. · School of Mechanical Engineering, Kyungpook National University, Daegu, 41566, Republic of Korea.

Abstract

Quadruped robots entering hospitals, care homes, and quiet offices must be context-appropriate not only in where they move but in how loudly they move: a legged robot's locomotion noise, dominated by foot-ground impacts, is itself a social variable. Prior social navigation respects human space but treats the robot as acoustically uniform, while quiet-locomotion methods reduce noise to an operator-specified, context-blind level. We present TACET, a context-appropriate acoustic-social navigation method that infers social context from the robot's egocentric view and decides both where it walks and how loudly, coupling a slow fine-tuned vision-language reasoner to a fast reactive controller through a single compact behavior token, <gait, speed, social_cost>. The same token conditions both a social costmap (where to go) and a quiet locomotion policy (how loudly to move), while a structured out-of-view memory keeps recently seen people in the reasoner's context after they leave the camera view. On a real quadruped, context-conditioned locomotion lowers locomotion noise by up to 9.3 dBA at matched speed, and across our scenarios the full method keeps personal-space compliance at 100% with low acoustic intrusion (<=2.9 dBA), jointly improving spatial and acoustic performance in the evaluated scenarios. The project page is available at https://rcilab.khu.ac.kr/tacet/.

Figures & tables

Fig. 1: Method paths on real runs of the two indoor scenarios: a focused person working at a table (left) and a standing group in conversation (right). Shading is the social costmap, rendered on the ground plane. TACET (red) swings wide and keeps its distance, whereas RPP (blue), MUTE (amber) and VLM-Social-Nav (magenta) take the short path past the people, and SFM (purple) enters the personal zone and passes closest to the seated person.

Fig. 2: Overview of TACET. The semantic reasoner (person tracker $\rightarrow$ social memory $\rightarrow$ VLM reasoner) emits the behavior token $\langle$ gait , speed , social_cost $\rangle$ ; the reactive controller (social costmap $\rightarrow$ path planner $\rightarrow$ velocity controller $\rightarrow$ locomotion policy) consumes it to produce socially aware, acoustically controlled motion.

Speed (m/s)	Method	MNL (dBA) $\,\downarrow$	PNL (dBA) $\,\downarrow$
0.5	Default Go2	$68.9\pm 2.8$	$72.5\pm 0.7$
	MUTE ( $\beta{=}1$ )	$67.8\pm 2.0$	$71.5\pm 0.5$
	Ours	$\mathbf{64.1\pm 1.3}$	$\mathbf{65.9\pm 0.8}$
0.8	Default Go2	$74.4\pm 2.4$	$77.2\pm 1.2$
	MUTE ( $\beta{=}1$ )	$71.7\pm 2.9$	$74.6\pm 0.7$
	Ours	$\mathbf{65.1\pm 2.1}$	$\mathbf{68.1\pm 1.6}$

TABLE I: Locomotion noise at matched speeds.

Scenario	Method	PSC $\,\uparrow$	MinDist $\,\uparrow$	$\Delta L\,\downarrow$	AID $\,\downarrow$
Focused (working)	RPP	95.6	$1.03\pm 0.12$	$8.4\pm 0.5$	66.3
	SFM	92.7	$0.41\pm 0.05$	$5.5\pm 0.3$	51.8
	MUTE	97.4	$0.91\pm 0.14$	$6.4\pm 0.3$	61.2
	VLM-SN	100.0	$1.35\pm 0.06$	$5.9\pm 0.2$	58.3
	Ours	100.0	$\mathbf{3.61\pm 0.17}$	$\mathbf{1.6\pm 0.2}$	0.0
Focused (headph.)	RPP	95.8	$1.06\pm 0.14$	$8.1\pm 0.6$	65.9

TABLE II: Social-navigation results across four scenarios.

Scenario	Time to goal (s)			Emitted token
Scenario	mean	best	Ours	gait	spd	cost
Focused (working)	31.8	21.2	36.8	quiet	low	4.0
Focused (headph.)	31.8	22.0	33.8	normal	low	4.0
Standing group	17.2	13.6	25.6	quiet	low	3.5
Moving person	16.7	14.8	21.8	quiet	normal	2.5
Empty corridor	16.2	14.4	14.4	normal	normal	1.5

TABLE III: Time to goal and emitted tokens by scenario.

Fig. 3: Per-question results of the blinded video study. Q1: safe and comfortable distance; Q2: contextual consideration; Q3: footstep-noise appropriateness; Q4: scenario-specific etiquette. Bars show mean scores on a 0–5 Likert scale; error bars show SEs. Higher is better.

Scenario	Variant	PSC $\,\uparrow$	MinDist $\,\uparrow$	$\Delta L\,\downarrow$	AID $\,\downarrow$
Focused (working)	Qwen2.5-VL-7B	100.0	$1.35\pm 0.10$	$4.5\pm 0.5$	31.8
	InternVL3.5-8B	100.0	$1.33\pm 0.12$	$4.9\pm 0.4$	35.1
	Ours (full)	100.0	$\mathbf{3.61\pm 0.17}$	$\mathbf{1.6\pm 0.2}$	0.0
	w/o soc. layer	96.5	$1.18\pm 0.11$	$4.3\pm 0.3$	29.3
	w/o trained pol.	100.0	$3.47\pm 0.19$	$7.4\pm 0.5$	63.3
	w/o soc. memory	100.0	$3.54\pm 0.24$	$3.9\pm 0.3$	18.7

TABLE IV: Ablations on two person-facing scenarios.

Robot Navigation

Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.

See all 536 papers

ACME: A Multi-Cultural, Multi-Embodiment Social-Navigation Dataset
Jul 24, 2026Shashank Rao Marpally, Allan Wang, Atharva Ghotavadekar +35Pedestrian Trajectory Prediction Robotic Data Collection
School of Computing, National University of Singapore · Miraikan - The National Museum of Emerging Science and Innovation · George Mason University +7
E-SocialNav: Efficient Socially Compliant Navigation with Language Models
Mar 21, 2026Ling Xiao, Daeun Song, Xuesu Xiao +1Social Robot Navigation
Hokkaido University, Sapporo, Japan · George Mason University, Virginia, United States · The University of Tokyo, Tokyo, Japan
Learning Social Robot Navigation By Sensing Human Legs
Jul 30, 2026Alberto Vaglio, Andrea Garulli, Antonio Giannitrapani +2RL for Robotics Robot Policy Learning
Department of Information Engineering and Mathematics, University of Siena, Siena, Italy · Uninettuno University, Rome, Italy

Jul 24, 2026J/K move · Enter open · S save

ACME: A Multi-Cultural, Multi-Embodiment Social-Navigation Dataset

Shashank Rao Marpally, Allan Wang, Atharva Ghotavadekar, Renato Alexandre Ribeiro+34

School of Computing, National University of Singapore · Miraikan - The National Museum of Emerging Science and Innovation · George Mason University+7

Pedestrian Trajectory Prediction Robotic Data Collection Social Robot Navigation Social Navigation

Open paper AskPDF arXiv

Understanding how robots and humans move in shared spaces is essential for designing effective social robot navigation policies and predicting human behavior. However, existing datasets often lack the diversity needed to capture differences in culture, geography, and human-robot interaction-factors that strongly shape appropriate social behavior. To address this gap, we introduce ACME: A Cross-cultural, Multi-Embodiment dataset for social navigation. A large-scale data collection effort across 8 sites in 5 countries, using 7 robot embodiments, ACME is a large and diverse multi-modal dataset aimed at advancing social navigation research, providing 29.35 hours of onboard robot data and 43.5 hours of overhead pedestrian tracking data. Unlike prior datasets, it focuses on capturing goal-driven social navigation behavior in complex social scenarios with explicit robot-crowd interaction through robot speech. To facilitate learning navigation policies and predicting pedestrian trajectories, ACME provides 3D and 2D scene features, odometry, interaction information, and human-annotated pedestrian trajectory labels. We make ACME easy to use by providing both human-readable data for each sensor modality as well as raw binary data. Our qualitative and quantitative analyses show that our dataset captures more challenging scenarios and a broader distribution of pedestrian behavior than previous datasets.

Abstract

Robot Navigation

Explore similar work

ACME: A Multi-Cultural, Multi-Embodiment Social-Navigation Dataset

E-SocialNav: Efficient Socially Compliant Navigation with Language Models

Learning Social Robot Navigation By Sensing Human Legs