A legged robot's locomotion noise is a social variable. TACET infers the social context of the scene and emits one compact behavior token, ⟨gait, speed, social_cost⟩, that sets both where the robot goes and how loudly it walks.
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.
A fine-tuned vision-language model reads the egocentric view and emits one behavior token at roughly 0.25 Hz — slow enough to reason, far outside the safety loop.
The same token scales an asymmetric proxemic costmap and conditions a quiet locomotion policy, so room and quietness are allocated by a single inferred signal.
A structured out-of-view memory carries recently seen people across field-of-view loss, each tagged with an age-decayed confidence.
A fine-tuned Qwen2.5-VL-7B maps the egocentric RGB image, a natural-language instruction, and a structured
social memory to a short chain-of-thought followed by the behavior token, with gait ∈
{quiet, normal, urgent}, speed ∈ {low, normal, high}, and social_cost a scalar
amplitude at a resolution of 0.5. Two categorical fields and one scalar make the output directly parseable,
in contrast to free-form language or continuous embeddings.
Supervision comes from auto-labeling: a large VLM produces structured ⟨thought, action token⟩ labels on 3,950 egocentric frames from MuSoHu, JRDB-Social, SCAND and InCrowd-VI, every one human-verified, combined with 18,042 COCO images auto-labeled into the same format and 909 SocialNav-SUB frames remapped onto the token format, for 34,751 training records over 22,901 distinct frames. On a held-out benchmark of 100 human-labeled SCAND frames, exact agreement rises from 30% / 26% / 14% for the base model to 95% / 96% / 69% after QLoRA fine-tuning. A query takes 2.6 s at the median, so tokens are refreshed at roughly 0.25 Hz and cached — slow reasoning never enters the safety loop.
A narrow camera field of view means people the robot was just reacting to leave the frame while remaining well inside its local planning neighborhood. People are detected with YOLOv8n-pose, localized by fusing depth with the point cloud, and maintained in the odometry frame under a hold / coast / refresh policy: a moving person is coasted at constant velocity, a stationary person held until a time-to-live expires. Each retained person becomes a structured record whose confidence starts from source quality and decays with observation age. Records enter the prompt as reasoning context, never as a rule-based override.
Each person induces an anisotropic Gaussian in their own frame, written into one layer of a layered costmap. What differs from prior proxemic costmaps is not the shape but where its parameters come from: in TACET the amplitude and width follow the emitted token, so the same decision that sets how quietly the robot walks also sets how much room it leaves. Collisions stay in a separate hard obstacle layer, so social cost modulates preference without compromising geometric safety.
Because quadruped locomotion noise is dominated by foot–ground impacts, whose acoustic energy grows with
foot velocity at contact, TACET reduces perceived noise through control rather than hardware. The
gait and speed tokens enter the policy as physical quantities rather than a
one-hot code: gait selects a contact schedule, a four-beat crawl for quiet and a trot with
decreasing duty factor for normal and urgent, and speed sets its period. One network therefore
covers every gait–speed pair the reasoner emits. A token-gated foot-impact penalty tightens as the gait
token moves from urgent to quiet at matched speed. Unlike an operator-set quiet factor, quietness is
allocated by context rather than preset.
social_cost token: circular (static), egg-shaped
elongated along travel (moving), and one enlarged field containing an interacting pair, protecting the
shared interaction space.
We first characterize the locomotion policy in isolation, with no people present, which separates the gait's acoustic contribution from that of any detour. Baselines are compared at matched actual speed taken from LiDAR-inertial odometry, so a method cannot appear quieter merely by moving more slowly.
| Speed (m/s) | Method | MNL (dBA) ↓ | PNL (dBA) ↓ |
|---|---|---|---|
| 0.5 | Default Go2 | 68.9 ± 2.8 | 72.5 ± 0.7 |
| MUTE (β=1) | 67.8 ± 2.0 | 71.5 ± 0.5 | |
| Ours | 64.1 ± 1.3 | 65.9 ± 0.8 | |
| 0.8 | Default Go2 | 74.4 ± 2.4 | 77.2 ± 1.2 |
| MUTE (β=1) | 71.7 ± 2.9 | 74.6 ± 0.7 | |
| Ours | 65.1 ± 2.1 | 68.1 ± 1.6 | |
| 1.1 | Default Go2 | 74.8 ± 1.9 | 76.9 ± 1.2 |
| MUTE (β=1) | 72.3 ± 2.5 | 76.6 ± 0.2 | |
| Ours | 70.6 ± 2.4 | 75.1 ± 0.6 |
Speed is the commanded speed. Mean ± standard deviation over ten trials per speed; lower is better.
Ours is the quietest at every matched speed, lowering mean noise level by up to 9.3 dBA over the default controller and 6.6 dBA over MUTE. Standing still with only its cooling fan running, the Go2 already registers about 60 dBA at the same microphone position — the floor these levels should be read against. The margin is smallest at 1.1 m/s, where impact noise dominates and little headroom remains.
Nothing in the instruction names a gait, a speed, or an amount of room. The token follows the scene, and the behavior follows the token — from a quiet crawl past someone working to an urgent trot past a siren light.
With a person in the robot's path
Evaluated with the full set of spatial and acoustic social metrics.
Someone absorbed in a stationary task in a quiet room. The robot must reach a nearby goal without acoustically intruding.
The same person, seat, posture and layout. Headphones are the only changed cue — and the gait token changes with them.
A small group converses. The robot must reach a goal beyond them without cutting through the shared interaction space.
A person walks through the space, where a higher ambient level masks part of the robot's own noise. The robot yields room as it passes.
With no one in the path
The social metrics do not apply here; what matters is that caution is not spent where it buys nothing.
No one present, so any detour or slowdown is pure overhead.
A large open space with people active far from the robot's path.
A flashing siren light in a dim corridor. Seeing it, the reasoner decides speed outweighs restraint.
PSC is the fraction of the run spent at least 1.2 m from every person; MinDist the closest approach; ΔL the A-weighted level the robot adds at the listener; AID the fraction of the run for which ΔL exceeds 5 dBA.
| Scenario | Method | PSC (%) ↑ | MinDist (m) ↑ | ΔL (dBA) ↓ | AID (%) ↓ |
|---|---|---|---|---|---|
| Focused (working) | RPP | 95.6 | 1.03 ± 0.12 | 8.4 ± 0.5 | 66.3 |
| SFM | 92.7 | 0.41 ± 0.05 | 5.5 ± 0.3 | 51.8 | |
| MUTE | 97.4 | 0.91 ± 0.14 | 6.4 ± 0.3 | 61.2 | |
| VLM-SN | 100.0 | 1.35 ± 0.06 | 5.9 ± 0.2 | 58.3 | |
| Ours | 100.0 | 3.61 ± 0.17 | 1.6 ± 0.2 | 0.0 | |
| Focused (headph.) | RPP | 95.8 | 1.06 ± 0.14 | 8.1 ± 0.6 | 65.9 |
| SFM | 92.2 | 0.39 ± 0.06 | 5.7 ± 0.2 | 52.3 | |
| MUTE | 97.1 | 0.89 ± 0.15 | 6.3 ± 0.3 | 60.1 | |
| VLM-SN | 100.0 | 1.35 ± 0.04 | 5.7 ± 0.2 | 57.9 | |
| Ours | 100.0 | 3.62 ± 0.19 | 2.9 ± 0.2 | 26.5 | |
| Standing group | RPP | 92.3 | 0.90 ± 0.19 | 12.1 ± 1.1 | 100.0 |
| SFM | 92.8 | 1.02 ± 0.02 | 7.8 ± 0.6 | 73.3 | |
| MUTE | 93.7 | 1.13 ± 0.05 | 4.9 ± 0.5 | 42.9 | |
| VLM-SN | 90.4 | 0.87 ± 0.16 | 7.3 ± 0.5 | 65.7 | |
| Ours | 100.0 | 2.25 ± 0.21 | 2.7 ± 0.3 | 22.3 | |
| Moving person | RPP | 87.7 | 0.59 ± 0.05 | 7.9 ± 0.7 | 83.2 |
| SFM | 90.2 | 0.75 ± 0.04 | 5.1 ± 0.5 | 51.3 | |
| MUTE | 98.9 | 1.21 ± 0.11 | 3.8 ± 0.3 | 24.7 | |
| VLM-SN | 94.1 | 0.69 ± 0.07 | 4.1 ± 0.4 | 29.4 | |
| Ours | 100.0 | 1.44 ± 0.12 | 1.7 ± 0.3 | 0.0 |
Means over ten trials, with standard deviations given for MinDist and ΔL. VLM-SN abbreviates VLM-Social-Nav.
Ours is the only method that keeps personal-space compliance at 100% in all four scenarios; VLM-Social-Nav reaches 100% with the focused person but not with the group or the moving person. Ours also keeps a far wider margin at the closest approach, with MinDist never below 1.44 m against at most 1.35 m for any baseline. All runs reached the goal without contacting a person, but not always by the robot's own doing. Three baseline conditions came within arm's reach of a person: SFM in both focused-person conditions, where it repeatedly stalled and scraped the wall (0.41 and 0.39 m), and RPP with the moving person, which stayed clear only because the person stepped aside (0.59 m).
The two focused-person conditions hold the scene fixed and change only the headphones. Methods that consume positions alone behave near-identically in both, as VLM-Social-Nav does (ΔL 5.9 vs. 5.7 dBA). Ours separates them, keeping the same room (MinDist 3.61 vs. 3.62 m) while relaxing the acoustic effort where the listener is less sensitive to it (ΔL 1.6 vs. 2.9 dBA, AID 0.0 vs. 26.5%). With the closest approach unchanged and only the gait token differing, the acoustic difference is consistent with the gait effect measured on the controlled traverse.
| Scenario | Baseline mean (s) | Baseline best (s) | Ours (s) |
|---|---|---|---|
| Focused (working) | 31.8 | 21.2 | 36.8 |
| Focused (headph.) | 31.8 | 22.0 | 33.8 |
| Standing group | 17.2 | 13.6 | 25.6 |
| Moving person | 16.7 | 14.8 | 21.8 |
| Empty corridor | 16.2 | 14.4 | 14.4 |
| Sports field | 18.3 | 14.4 | 14.0 |
| Emergency | 15.4 | 14.2 | 11.2 |
Mean time to goal over ten trials; the baseline columns summarize RPP, SFM, MUTE and VLM-Social-Nav. Bold marks the fastest entry in each row.
Where people need space or quiet, TACET spends at most 1.5× the mean baseline time. Where they do not, it matches the fastest baseline on an empty corridor and is the fastest method outright on the open field and with the siren light.
We ran a blinded video study with 15 observers to complement the objective results. Each observer rated the 20 short clips formed by the four person-facing scenarios and five methods (RPP, SFM, MUTE, VLM-Social-Nav, and TACET), presented in randomized order and without method labels. After every clip, the observer rated Q1–Q4 on a 0–5 agreement scale. All sessions used the same headphones and playback level so that relative footstep-noise differences were preserved.
0: strongly disagree · 1: disagree · 2: somewhat disagree · 3: somewhat agree · 4: agree · 5: strongly agree.
TACET receives the highest agreement in all sixteen scenario–question cells, averaging 4.54 against 2.23 for MUTE, 1.92 for VLM-Social-Nav, 1.07 for SFM and 0.78 for RPP, and its standard errors are the smallest (0.14 on average against 0.29 for VLM-Social-Nav), so observers agreed about it more than about any baseline. The margin narrows only with the moving person, where MUTE reaches 3.33–4.07 — the same scenario in which MUTE is the strongest baseline on the objective metrics, so its quiet gait alone recovers much of what observers reward.
One component is removed at a time on two representative person-facing scenarios. Separately, the top block of each scenario keeps the architecture fixed and swaps only the reasoner that emits the token, isolating the contribution of fine-tuning beyond prompt engineering.
Clearance is what the social layer buys; quietness is what the trained policy buys. Removing either one leaves the other intact. The clip then swaps the reasoner for a prompted base VLM, which keeps the architecture fixed and loses both.
People pass in and out of the camera view as the robot walks, so this is where the social memory earns its place. The reasoner swap that follows barely changes the route here, yet the robot turns audible above ambient for a third of the run.
| Scenario | Variant | PSC (%) ↑ | MinDist (m) ↑ | ΔL (dBA) ↓ | AID (%) ↓ |
|---|---|---|---|---|---|
| Focused (working) | Qwen2.5-VL-7B | 100.0 | 1.35 ± 0.10 | 4.5 ± 0.5 | 31.8 |
| InternVL3.5-8B | 100.0 | 1.33 ± 0.12 | 4.9 ± 0.4 | 35.1 | |
| Ours (full) | 100.0 | 3.61 ± 0.17 | 1.6 ± 0.2 | 0.0 | |
| w/o social layer | 96.5 | 1.18 ± 0.11 | 4.3 ± 0.3 | 29.3 | |
| w/o trained policy | 100.0 | 3.47 ± 0.19 | 7.4 ± 0.5 | 63.3 | |
| w/o social memory | 100.0 | 3.54 ± 0.24 | 3.9 ± 0.3 | 18.7 | |
| Moving person | Qwen2.5-VL-7B | 100.0 | 1.38 ± 0.11 | 3.5 ± 0.5 | 36.1 |
| InternVL3.5-8B | 100.0 | 1.33 ± 0.12 | 3.8 ± 0.7 | 39.3 | |
| Ours (full) | 100.0 | 1.44 ± 0.12 | 1.7 ± 0.3 | 0.0 | |
| w/o social layer | 94.7 | 0.89 ± 0.12 | 3.8 ± 0.3 | 37.4 | |
| w/o trained policy | 100.0 | 1.39 ± 0.13 | 4.2 ± 0.5 | 41.5 | |
| w/o social memory | 100.0 | 1.42 ± 0.09 | 2.6 ± 0.4 | 19.3 |
Means over ten trials, with standard deviations given for MinDist and ΔL. Rows above each rule swap only the reasoner that emits the token; rows below remove one component.
Distance and gait are separable. Removing the trained policy keeps the closest approach within one standard deviation of the full method, yet degrades the acoustic terms the most of any variant, by 5.8 dBA with the focused person and 2.5 dBA with the moving one, supporting an acoustic contribution from the gait policy beyond that of the detour. Removing the social layer costs clearance instead, collapsing MinDist from 3.61 to 1.18 m and from 1.44 to 0.89 m, both inside the 1.2 m zone the full method never enters. Removing social memory barely alters the route, yet the robot becomes audible above ambient for 18.7% and 19.3% of the run: without it, a person who briefly leaves the camera view is no longer carried in the reasoner's context.
Swapping the reasoner degrades both axes. With the focused person, a prompted Qwen2.5-VL-7B keeps only 1.35 m of clearance against 3.61 m for the full method, close to what removing the social layer gives — and raises ΔL to 4.5 dBA with AID at 31.8%; InternVL3.5-8B behaves the same way (1.33 m, 4.9 dBA, 35.1%). With the moving person the route changes far less (1.38 and 1.33 m against 1.44 m), yet the robot is audible above ambient for 36.1% and 39.3% of the run against 0.0% for the full method. Since the architecture is unchanged, the loss follows from the token interface itself rather than from any component.
Each tracked person contributes a bounded, anisotropic social field that changes the Nav2 path preference;
the obstacle layer remains responsible for collision avoidance. The parameters below are the ones the paper
leaves symbolic. The social_cost token, filtered and written s, is normalised into
ρ, which drives both the field amplitude A and its width σ:
ρ = max(ρfloor, min(1, s / smax))
A = A0 + (Amax − A0) ρ
σ = σ0(1 + 0.58 ρ) · Mgait · Mspeed · kshape
σfront = σ(1 + 0.45 m) for m > m0, else σside = σrear = σ
A floor ρfloor keeps a minimum field even when the token is small, while the hard obstacle
core below guarantees geometric clearance independently. Above the floor, the social_cost token,
the gait multiplier and the speed multiplier all contribute to the field width.
The reward combines velocity tracking with acoustic, gait-style, symmetry and regularization
terms. The acoustic terms are token-gated: they scale with the quietness of the
gait token, so at a matched speed the quiet token lands softer than the normal or
urgent token — from the same network. Falls are handled by episode termination rather than
a collision penalty.
Symmetry deserves its own mention. A yaw-rate penalty on straight commands, a left–right joint-mirror term and a contact-force balance term together suppress the lateral drift that a quiet gait otherwise inherits, keeping a straight command straight without giving up the softer footfall.
| Group | Term | What it shapes | Weight |
|---|---|---|---|
| Task | Linear velocity tracking | Exponential on the squared command error | 1.0 |
| Angular velocity tracking | Same form on yaw rate | 2.5 | |
| Move shortfall | Anti-stall: progress falling short of the commanded speed | −0.9 | |
| Acoustic | Token-scaled foot velocity | Squared vertical foot speed, capped | −0.55 |
| Phase drop / raise | Soft landing near touchdown, small bonus for lifting | 1.0 | |
| Peak contact-force cap | Contact force above 120 N | −1.0 | |
| Contact loading rate | Force jumping by more than 10 N in one step | −2×10−3 | |
| Gait style | Contact-schedule tracking | Stance leg loaded, swing leg unloaded, on the token's phase | 0.8 |
| Foot clearance | Swing height up to 28 mm | 2.0 | |
| Foot lift | Secondary clearance term up to 35 mm | 0.3 | |
| Feet air time | Air time beyond 0.5 s at touchdown | 0.125 | |
| Stance width | Lateral foot offset up to 0.17 m | 2.5 | |
| Symmetry | Straight-yaw penalty | Yaw rate while the command is straight | −30.0 |
| L/R joint mirror | Thigh and calf mismatch between left and right | −2.0 | |
| L/R force balance | Contact-force mismatch between left and right | −2×10−4 | |
| Posture | Fore–aft pitch | Pitch away from flat | −8.0 |
| Orientation | Roll and pitch of the gravity vector | −0.08 | |
| Vertical velocity | Base bobbing | −5.0 | |
| Angular velocity (xy) | Roll and pitch rate | −0.3 | |
| Smoothness | Action bound | Actions driven past their working range | −0.1 |
| Action rate | Change between consecutive actions | −0.02 | |
| Joint torque | Squared torque | −2×10−4 | |
| Joint power | Torque times joint velocity | −2×10−5 | |
| Joint acceleration | Squared joint acceleration | −2.5×10−7 |
Positive weights reward, negative weights penalize. Phase drop / raise is a single signed term whose net effect is a soft-landing penalty.
@inproceedings{anonymous2027tacet,
title = {TACET: Context-Appropriate Acoustic-Social Navigation for Quadrupeds},
author = {Anonymous Authors},
booktitle = {Proc. IEEE Int. Conf. Robotics and Automation (ICRA)},
note = {Submitted},
year = {2027}
}