Walk with Me: Mobility Interaction Research

Updated 4 July 2026

Walk with Me is a broad research paradigm that leverages walking as a central medium for human-centric navigation, assistive guidance, and ethnographic study.
It integrates diverse methodologies such as map-free navigation, gait entrainment, and augmented reality to coordinate locomotion, perception, and social context.
Recurrent findings highlight that synchronous motion, temporal precision, and socially compliant cues enhance performance despite challenges like sensor noise and personalization gaps.

“Walk with Me” refers to a broad research agenda in which walking is not merely a background activity but the primary substrate of interaction, control, assistance, observation, or shared experience. In current work, the term spans map-free long-horizon outdoor assistance for mobile robots, public-pathway ethnography that physically accompanies delivery robots, socially assistive walking companions, blind-walking guidance, gait-training avatars, virtual walking experiences, mobile coaching systems, and remote cadence-sharing interfaces (Zhang et al., 29 Apr 2026, Cheon et al., 18 Feb 2026, Garcia et al., 2019, Wijesooriya et al., 2023, Baldi et al., 2020). Across these domains, the central problem is how to coordinate locomotion, perception, timing, and social context over extended trajectories rather than isolated commands.

1. Conceptual scope and theoretical foundations

Within human–robot interaction and adjacent fields, “walk with me” has no single canonical meaning. In one line of work, it denotes human-centric outdoor navigation: a robot receives a high-level natural-language intention such as “I want to go for a walk” and must ground that intention into destinations, waypoints, and socially compliant movement in open-world outdoor settings (Zhang et al., 29 Apr 2026). In another, it denotes public-field ethnography: researchers physically accompany commercial delivery robots in city streets to study how robot movement is co-constituted by pedestrians, infrastructures, and temporal rhythms (Cheon et al., 18 Feb 2026). In others, it denotes walking companionship in rehabilitation, blind-walking assistance, gait synchronization, XR embodiment, or conversational coaching (Garcia et al., 2019, Wijesooriya et al., 2023).

The methodological foundations are correspondingly heterogeneous. The Walk-Along with Robots methodology explicitly draws on the “mobilities turn” in urban studies and geography associated with Sheller and Urry and with Cresswell, as well as mobile methods such as the go-along of Kusenbach and the walking interviews of Jones et al. It also adopts a more-than-human perspective, treating robots as active “interlocutors” rather than passive artifacts (Cheon et al., 18 Feb 2026). By contrast, outdoor assistance frameworks emphasize semantic grounding, GPS context, public map APIs, and hierarchical VLM/VLA dispatch for safety-critical navigation (Zhang et al., 29 Apr 2026).

A common misconception is that “walk with me” denotes only a robot following a person. The literature is broader. It includes side-by-side gait entrainment with an augmented-reality avatar, remote peer-to-peer cadence sharing via ankle haptics, voice-guided and gesture-guided control of a quadruped while walking alongside it, and even real-time musical adaptation to footfalls (Wijesooriya et al., 2023, Baldi et al., 2020, Zhang et al., 2024, James, 1 Jun 2025). This suggests that the phrase designates a family of movement-centered interaction paradigms rather than a single technical architecture.

2. Walk-along ethnography in public pathways

The most explicit methodological formulation is Walk-Along with Robots (WawR), proposed for studying autonomous delivery robots that operate beyond the control of researchers in crowded public pathways (Cheon et al., 18 Feb 2026). WawR begins with preparatory desk research: service areas are mapped using delivery-service apps, partner stores and pick-up points are bookmarked on a local mapping tool, and likely routes, traffic controls, pedestrian densities, and hard-to-see obstacles are identified. It then incorporates embodied autoethnography, including placing real orders, following robots from docking to delivery, and comparing robot pace with local walking norms.

Fieldwork is organized around “pre-robot operation waiting areas” such as benches or café seating near known docking stations. Once a robot departs, researchers follow at varying distances: close for fine-grained interactional detail and farther back for situational overview. The daily routine described in the study includes arriving one hour before the robots’ operating window of 9 AM–4 PM, revisiting previous routes, taking position at a waiting area with clear sightlines, beginning continuous walk-along when the robot launches, exploiting natural pauses for short intercept interviews, and returning to the docking vantage after each delivery cycle (Cheon et al., 18 Feb 2026).

The observational log is intentionally multi-axial. It records routes, timestamps, encountered people, environmental cues such as horn versus voice prompts, obstacle types such as bollards, drainage grates, and banners, and researcher autoethnographic reflections. Team-based ethnography is central: one or two researchers shadow continuously, while another conducts intercept interviews during pauses such as long traffic signals. The methodology is evaluated by Data Completeness, Contextual Linkage, Triangulation & Team Coordination, Cultural Sensitivity & Naturalism, and Validation through Repetition. The account also sketches a three-axis framework—People, Places, and Time—and notes that a formalization could take the form $\mathrm{ObservationalCoverage} = f(\mathrm{People}, \mathrm{Places}, \mathrm{Time})$ (Cheon et al., 18 Feb 2026).

WawR’s distinctive empirical findings are strongly situated. Delivery robots “consistently emitted auditory prompts—either a car-horn sound or ‘Hello, I’m delivering. Let me pass by’—and rarely re-routed to avoid humans,” while pedestrians habitually yielded, producing what the paper describes as asymmetrical right-of-way. In a fast-paced mall precinct, pedestrians were often fixated on phones and near misses occurred at knee height; in a residential/business mix, robots attracted curiosity, touches, and selfies. People maintained strict personal distances from each other yet sometimes patted robots or opened food compartments without permission, suggesting unstable social norms around robot handling. Full-day shadowing also revealed lunchtime clustering in restaurant zones, mid-afternoon lulls, and shifts in public attitudes from nuisance to neighborhood fixture (Cheon et al., 18 Feb 2026).

In outdoor autonomy, “Walk with Me” has been formalized as a map-free framework for long-horizon social navigation from high-level human instructions (Zhang et al., 29 Apr 2026). The architecture is hierarchical. A High-Level Semantic Planner takes a user instruction $l$ , GPS context, and nearby place candidates $C=\{c_i\}$ returned by a public map API, then selects a destination by

$c^*=\mathrm{arg\,max}_{c_i\in C}\,\mathrm{score}_{\mathrm{VLM}}(l,c_i).$

Once $c^*$ is chosen, the system retrieves a walking route and resamples it into geo-referenced waypoints with semantic tags such as “crosswalk ahead.” During execution, it forms the observation

$o_t=(I_t,s_t,H_t,w_t^{\mathrm{goal}})$

and queries a high-level VLM to decide whether the current situation is routine or complex, and whether it is safe to proceed or necessary to stop-and-wait (Zhang et al., 29 Apr 2026).

The dispatch rule is explicitly safety-aware. If the decision is “proceed” and the confidence exceeds a threshold $\alpha$ , a Low-Level Vision-Language-Action policy generates a short-horizon body-frame trajectory $\hat W_t^{\mathrm{body}}$ . Otherwise the robot stops, emits zero velocity, and re-queries after a pause. The implementation reported no end-to-end retraining of either the high-level VLM or the low-level VLA; instead it used prompt engineering and off-the-shelf components such as Claude Sonnet, Gemini-3 Flash, GPT-5, Qwen3-VL-8B, MiMo-Embodied, RoboBrain 2.0, SocialNav, CityWalker, NoMaD, ViNT, and GNM (Zhang et al., 29 Apr 2026).

The real-world evaluation used an Athena 2.0 Pro AGV with RGB-D camera, GPS, SLAM, and a remote inference server. Across 20 total trials—five per scenario for last-mile delivery and blind guidance tasks—the overall Success Rate was 60% (12/20). The reported scenario-level rates were 70% for last-mile delivery and 50% for blind guidance. In ablations, MiMo-Embodied was the best high-level VLM at 60% SR, while SocialNav was the best low-level VLA at 60%, followed by CityWalker at 50%, NoMaD at 30%, and ViNT/GNM at 20% (Zhang et al., 29 Apr 2026).

A related low-level policy is CityWalker, which learns urban navigation from approximately 2,000 h of first-person city walking footage plus auxiliary driving videos (Liu et al., 2024). Its pipeline uses DPVO for relative-pose extraction, normalizes per-clip motion, encodes RGB with frozen DINOv2 (ViT-B/14), embeds past motions and a target waypoint, and predicts future displacements with a 16-layer, 8-head transformer. At runtime, the policy consumes live camera frames, GPS position, and a next navigation-tool waypoint, then sends the first predicted displacement to a low-level PD controller at approximately 10 Hz. On 9 h of teleoperation test data, fine-tuned CityWalker reported MAOE≈15.2° and Arrival≈88%, versus MAOE≈16.5° and Arrival≈71% for the best fine-tuned ViNT baseline. In real-world quadruped trials, CityWalker achieved 77.3% overall success, compared with 57% for fine-tuned ViNT and 42.9% for zero-shot NoMaD (Liu et al., 2024).

A second misconception concerns the phrase map-free. In these systems, map-free does not imply the absence of external priors. The reported architectures still rely on GPS context, public map APIs, and navigation-tool waypoints for destination grounding or sub-goal definition (Zhang et al., 29 Apr 2026, Liu et al., 2024).

4. Control modalities, alignment, and shared locomotor timing

One important strand of “walk with me” research studies how a human directly controls or synchronizes with a moving robot. In a 2×2 fully within-subject design (N=218) using the quadruped Spot, participants issued either voice commands or mid-air hand gestures while either standing still or walking alongside the robot (Zhang et al., 2024). The task required guiding Spot through a 14-node, 90°-turn trajectory using three commands: Walk Forward (1 m), Rotate Left (90°), and Rotate Right (90°). The study’s alignment metric was

$\mathrm{alignment}\% = 100\% \cdot \frac{\text{number of time samples with } |\theta_{\mathrm{human}}-\theta_{\mathrm{robot}}|\le 45^\circ}{\text{total samples}}.$

Voice control was faster overall than gesture control (5.50 ± 0.60 s vs. 5.96 ± 0.58 s inter-command time), had higher self-reported command detection (4.60 ± 0.54 vs. 4.22 ± 0.59), and was preferred by 71% of participants, while 29% preferred gestures. Walking improved mapping intuitiveness for both modalities, and gesture control while standing produced elevated left/right reversals (GS 0.21 vs. GW 0.03) (Zhang et al., 2024).

The older c-Walker literature approached the same problem through path-following assistance rather than robot teleoperation. It compared mechanical steering, haptic cues, Left/Right acoustic cues, and binaural guidance in a within-subjects study with N=13 following three 10 m virtual corridors (Moro et al., 2016). The best path-following accuracy came from mechanical steering, with average lateral error 18.1 cm, compared with 39.6 cm for acoustic, 74.0 cm for haptic, and 62.4 cm for binaural guidance. Mechanical steering was also faster than haptic guidance, but participants described it as potentially coercive. Haptic bracelets were judged more socially acceptable than headphones and left vision and audition free, whereas acoustic cues drew complaints about headphone bulk and occlusion of environmental sounds (Moro et al., 2016).

A third line of work studies gait entrainment rather than directional control. Wearable Haptics for Remote Social Walking used ankle bands with vibro-motors and a heel-mounted force-sensing resistor to stream gait cadence between walkers over smartphone and server connections (Baldi et al., 2020). The reported end-to-end latency was <100 ms, with heel-strike sensing at 100 Hz and vibro-motor lag of approximately 20 ms. Across a four-stage validation with 50 total participants, median alignment reached 99.3% for an artificial constant reference, 98.8% and 99.0% for artificial variable faster/slower references, 96.8 ± 1.8% for a human leader, and 94.5 ± 4.1% for peer-to-peer coupling. User-reported ratings included 5.7 ± 1.1 for “Feels human-like?” and 5.9 ± 1.1 for “Perceived presence?” on 1–7 Likert scales (Baldi et al., 2020).

Taken together, these studies show that walking itself can function as a control resource. When people can reposition relative to the robot, alignment becomes more egocentric; when the guidance signal is embedded in cadence, synchronization can occur within natural gait variability (Zhang et al., 2024, Baldi et al., 2020). This suggests that locomotor coupling is not merely an output variable but part of the interface.

In rehabilitation robotics, “walk with me” often denotes a physically proximate companion whose motion adapts to the user’s pace and intention. A feasibility study based on Pepper implemented a sensorless compliance approach in which external interaction force is inferred from discrepancies between nominal and actual joint torques:

$\tau_{\text{ext}}=\tau_{\text{total}}-\tau_f-\tau_m,$

with directional intent derived from

$l$ 0

These quantities are combined into a motion-intention vector $l$ 1, while obstacle-aware compliance is modulated by

$l$ 2

The rehabilitation-centre trial involved four patients—multiple sclerosis, Parkinson’s disease, left-side hemiplegia, and recent knee arthroplasty—each in a roughly 30-minute session. Robot speed was capped at 0.35 m/s and maximum acceleration at 0.3 m/s². All four patients learned the push/pull interface in under five minutes, preferred in-place rotation over lateral translation, and revealed the need for personalized contact zones and a no-touch follower mode (Garcia et al., 2019).

For blind and low-vision mobility, quadruped guide-dog robots have been developed with explicit attention to acoustic and physical comfort. The Unitree Go1 controller in “Human-Centered Development of Guide Dog Robots” combines a gait and reference generator, real-time nonlinear MPC solved by SQP and OSQP, a Whole-Body Impulse Controller, and perception from an Intel RealSense D435 with elevation mapping and stair detection (Yu et al., 17 May 2025). Compared with the default locomotion controller, the proposed controller yielded approximately 10 dB lower sound pressure, around 50 dBA versus around 60 dBA, while maintaining human walking speed. At 1.0 m/s, reported step length was approximately 0.32 m versus approximately 0.12 m for the default. In challenging-terrain tests on a soapy whiteboard, random wooden blocks, and a cardboard pile, it succeeded 10/10 times. In mixed-methods indoor experiments with four blind or legally-blind guide-dog handlers, the controller was rated quieter and more satisfactory, had SUS >80 versus approximately 60 for the default controller, and was described as “natural,” “soft,” and “comfortable” (Yu et al., 17 May 2025).

Blind-walking assistance has also been formulated as a streaming VLM problem. WalkVLM introduced a benchmark of approximately 12,000 video–annotation pairs, about 10 h of footage across 10 cities, with static tags, free-text scene descriptions, QA pairs, reminders, and temporal labels (Yuan et al., 2024). The model uses a three-stage hierarchical planner—static attribute extraction, scene summarization, and subgoal generation—plus a temporal gating score

$l$ 3

to decide whether to emit a reminder. On the held-out benchmark, WalkVLM reported 84.5% Guidance Accuracy, 1.2 alerts/min redundancy rate, 115 ms inference time, and 0.81 temporal-redundancy F1, outperforming Yi-VL (75.6%, 2.5, 145 ms), MiniCPM-V2.6 (78.9%, 2.1, 130 ms), and GPT-4o (80.3%, 1.9, 210 ms) (Yuan et al., 2024).

These works indicate that assistive walking systems now span a continuum from compliant physical support to perception-driven advisory guidance. Quiet locomotion, stable balance, stop-and-wait behavior, and concise reminders are all treated as accessibility requirements rather than auxiliary features (Garcia et al., 2019, Yu et al., 17 May 2025, Yuan et al., 2024).

6. Virtual, augmented, mobile, and mediated walking companions

Not all “walk with me” systems involve a physical robot. ARWalker is an augmented-reality application that projects a healthy-gait avatar in front of the user and relies on human–avatar coupling, including complexity matching and the Proteus effect, to induce synchronization (Wijesooriya et al., 2023). Implemented on Microsoft HoloLens 2 with Unity 2021 LTS, MRTK, and an Xsens motion-capture suit for baseline animation, it infers user walking speed from head pose and triggers avatar walking when $l$ 4 exceeds approximately 0.1 m/s. The avatar’s stride-time sequence can be isochronous, white noise, or pink noise (1/f), with playback speed set cycle-by-cycle by $l$ 5. On HoloLens 2, the prototype ran at 58–60 fps with end-to-end latency under 50 ms (Wijesooriya et al., 2023).

Virtual Steps examined the phenomenology of walking for a lifelong wheelchair user in VR over 9 days (Taheri et al., 2024). The participant settled on a nominal speed of 1.0 m/s, with acceleration and deceleration in three 0.25 m/s increments. Turning used continuous rotation at 45° per second, capped at 30° displacement per input, after 45° instant snap-turning caused disorientation. Head oscillation followed Lécuyer et al. with amplitude vector $l$ 6, stride length $l$ 7, and period $l$ 8, with footstep sounds synchronized every 0.8 s. Emotional Engagement accounted for 48% of coded diary segments, Agency/Control for 17%, and Cybersickness for 2%. The study further reported that command latency above 0.3 s and head-bob/footstep phase shifts above 100 ms disrupted agency and immersion (Taheri et al., 2024).

Mobile coaching systems move the paradigm toward conversational accompaniment. SmartWalkCoach uses a Bridge Agent, GeographyAgent, AccompanyAgent, and SummaryAgent to support route curation, just-in-time prompting, and post-walk reflection (Zhang et al., 14 May 2026). The AccompanyAgent runs in lightweight ticks of approximately 1 Hz, segments routes into virtual geofences, and triggers on either a segment boundary or a fatigue event defined by a 20% pace drop relative to the previous segment. In an in-the-wild AB/BA crossover study (N=12) comparing Information-only against Information+Motivation, linear mixed models showed that the motivational condition improved Positive Feelings with $l$ 9, $C=\{c_i\}$ 0, Cohen’s $C=\{c_i\}$ 1, and improved User Experience with $C=\{c_i\}$ 2, $C=\{c_i\}$ 3, $C=\{c_i\}$ 4, with no evidence of carryover (Zhang et al., 14 May 2026).

The same broad ecology includes systems where walking is measured, simulated, or repurposed rather than accompanied in the classical sense. Walk4Me turns an iPhone strapped to the trunk into a telehealth gait-assessment system sampling the accelerometer at 50 Hz and performing feature extraction plus classical and deep learning. In Duchenne muscular dystrophy and post-stroke cohorts, the best models reported 100.0% classification accuracy, 100.0% sensitivity, and 100.0% specificity under leave-one-subject-out cross-validation (Ramli et al., 2023). TouchWalker uses two-finger touchscreen “walking” to drive full-body avatar locomotion with a MoE-GRU model at 28–32 FPS on a Galaxy Tab S7+; in a within-subject study with N=14, it improved embodiment, enjoyment, and immersion relative to a virtual joystick, while performing worse in some fast-paced obstacle scenarios (Park et al., 11 Nov 2025). Iola Walker detects footfalls from a shoe-mounted IMU at 200 Hz using a Conv–LSTM on Android, with approximately 7 ms inference per 0.5 s window and a current end-to-ding latency of about 500 ms; on held-out data, the best reported F1 was 0.326 at ±15 samples (±75 ms) tolerance (James, 1 Jun 2025).

A common misconception is that virtual or mobile companions are necessarily less technical than physical walking robots. The reported systems involve explicit gait-noise models, synchronization constraints, temporal gating, recurrent or transformer architectures, and controlled user studies (Wijesooriya et al., 2023, Taheri et al., 2024, Zhang et al., 14 May 2026, Park et al., 11 Nov 2025).

7. Evaluation regimes, recurrent findings, and open problems

Across the literature, evaluation occurs at several distinct levels. One level concerns trajectory and navigation performance, using metrics such as Success Rate, AOE/MAOE, arrival success, lateral error, completion time, path length, and velocity ripple (Zhang et al., 29 Apr 2026, Liu et al., 2024, Moro et al., 2016, Yu et al., 17 May 2025). A second concerns interaction and alignment, using inter-command time, mistake counts, command detection, intuitiveness, alignment percentage, workload, satisfaction, SUS, and presence (Zhang et al., 2024, Baldi et al., 2020, Yu et al., 17 May 2025). A third concerns field methodology, using completeness across People, Places, and Time; contextual linkage; triangulation; naturalism; and repetition (Cheon et al., 18 Feb 2026). A fourth concerns streaming assistance and digital mediation, using reminder redundancy, inference time, F1, latency, and composite affect or UX scores (Yuan et al., 2024, Zhang et al., 14 May 2026, James, 1 Jun 2025).

Several recurrent findings appear across otherwise disparate systems. First, allowing a person to walk with the system, rather than remain stationary, often improves alignment or intuitiveness. This is explicit in the Spot study, where voice plus walking was the most favored condition and walking increased alignment for both voice and gestures (Zhang et al., 2024). Second, timing quality is repeatedly treated as central: WawR stresses temporal rhythms in public pathways, WalkVLM penalizes reminder redundancy, SmartWalkCoach suppresses prompts during high-load moments, and Virtual Steps identifies tight temporal synchrony between head bob and footstep audio as critical for agency (Cheon et al., 18 Feb 2026, Yuan et al., 2024, Zhang et al., 14 May 2026, Taheri et al., 2024). Third, social compliance is broader than collision avoidance. It includes asymmetrical right-of-way, stop-and-wait at crossings, quiet locomotion for blind users, supportive relational expression, and public acceptability of cueing modalities (Cheon et al., 18 Feb 2026, Zhang et al., 29 Apr 2026, Yu et al., 17 May 2025, Moro et al., 2016, Zhang et al., 14 May 2026).

The limitations are equally recurrent. Outdoor map-free systems remain sensitive to GPS noise, SLAM noise, and the coverage or freshness of public map services, and CityWalker’s generalization to irregular old-town layouts was identified as needing validation (Zhang et al., 29 Apr 2026, Liu et al., 2024). Public-space ethnography requires discretion and repeated observation but remains labor-intensive (Cheon et al., 18 Feb 2026). ARWalker reports systems performance but states that clinical efficacy results are for future work rather than the current publication (Wijesooriya et al., 2023). Guide-dog robots still face residual mechanical noise, fan noise, stair-geometry limitations, and small participant pools (Yu et al., 17 May 2025). Mobile coaching systems identify expression, timing, and frequency as unresolved personalization problems, and WalkVLM notes bias in obstacle recognition and limited geographic coverage (Zhang et al., 14 May 2026, Yuan et al., 2024).

Taken together, the literature suggests that “walk with me” is best understood as an integration problem. Successful systems must simultaneously manage locomotor dynamics, semantic intent grounding, perception, temporal coordination, accessibility, and social norms. None of the surveyed approaches solves all of these dimensions at once, but the field has already moved beyond isolated navigation or interface studies toward end-to-end walking ecologies in which people, robots, avatars, sensors, and urban environments are analyzed as coupled moving systems.