The reviewed record of science sign in
Pith

arxiv: 2607.06383 · v1 · pith:5V4OO5S2 · submitted 2026-07-07 · cs.RO

Towards Real-World Applications with an Autonomous Powered Wheelchair

Simone Arreghini , Alessandro Giusti , Alex Bordini , Enrico Ferrara , Giovanni Fulgoni , Antonio Paolillo This is my paper

Reviewed by Pith2026-07-08 07:34 UTCglm-5.2pith:5V4OO5S2open to challenge →

classification cs.RO
keywords wheelchairautonomousassistivemobilitynavigationpoweredperceptionreal-world
0
0 comments X

The pith

Self-Balancing Wheelchair Drives Itself to You

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper demonstrates that a commercial self-balancing wheelchair (Genny Zero) can be augmented with off-the-shelf sensing and a standard robotic navigation stack to perform autonomous assistive behaviors. The key integration is a software-control interface that translates standard velocity commands into the pitch and steering inputs the self-balancing platform natively accepts, allowing the navigation pipeline to treat the wheelchair as a planar differential-drive robot while the platform's internal controller handles balance. Using LiDAR for localization and mapping, an RGB-D camera for body tracking and gesture recognition, and the ROS 2 Nav2 framework for planning and control, the authors implement and demonstrate two scenarios: people-following (the wheelchair tracks a walking user through doorways and corridors) and hailing (a user summons the wheelchair from several meters away and it navigates autonomously to their position). The experiments were conducted indoors at low speeds with the wheelchair unoccupied. The authors position this as a proof of concept that bridges a real assistive mobility platform with autonomous robotics, and they identify the main integration challenges: the quasi-static navigation approximation causes stop-and-go oscillations that interact with the self-balancing dynamics, frontal-only sensing limits coverage, and the current torque limit prevents operation with a seated user.

Core claim

The central finding is that the self-balancing wheelchair's internal pitch-based control can be driven by an outer PID velocity loop, allowing standard planar navigation software to command the platform without directly modeling or accessing the balancing dynamics. This abstraction is sufficient for low-speed, indoor, proof-of-concept demonstrations of people-following and hailing, but it breaks down when command jitter from the navigation layer introduces stop-and-go behavior that couples with the platform's self-balancing dynamics, producing oscillations that can destabilize perception and tracking.

What carries the argument

Genny Zero self-balancing wheelchair; software-control mode converting velocity commands to pitch and steering; ROS 2 Nav2 navigation stack; AMCL localization with differential-drive motion model in the base_footprint frame; LiDAR-based virtual 2D scan; RGB-D body tracking for gesture recognition (hailing, follow-me, stop); Regulated Pure Pursuit controller

If this is right

  • If the velocity-to-pitch interface proves robust at higher speeds and with a seated user, autonomous assistive features (hailing, people-following, obstacle-aware navigation) could be deployed on existing self-balancing wheelchairs without redesigning the low-level balancing controller.
  • The stop-and-go oscillation problem suggests that navigation planners for self-balancing platforms must produce smooth, continuous velocity profiles rather than the discrete replan-and-stop patterns typical of planar mobile-robot stacks.
  • Shared-control driver-assistance functions (forward obstacle awareness, speed modulation, corridor centering) could be built on the same software-control interface, potentially offering a lower-risk path to real-world deployment than full autonomy.
  • Extending perception to 360-degree coverage while preserving the wheelchair's form factor and usability remains an unsolved engineering challenge specific to assistive platforms.

Load-bearing premise

The navigation pipeline treats the wheelchair as a planar differential-drive robot and ignores the pitch dynamics of the self-balancing platform entirely. This quasi-static approximation is load-bearing because all demonstrated behaviors depend on it, and the paper itself shows that it causes oscillations when the planner sends intermittent stop-and-go commands that couple with the balancing controller.

What would settle it

If pitch coupling between navigation commands and the balancing controller cannot be neglected at speeds above 1 m/s or when a user is seated on the platform, the velocity-to-pitch control interface would be insufficient for any operationally useful autonomous behavior, and the entire integration approach would need to be replaced with a controller that jointly models navigation and balance dynamics.

Figures

Figures reproduced from arXiv: 2607.06383 by Alessandro Giusti, Alex Bordini, Antonio Paolillo, Enrico Ferrara, Giovanni Fulgoni, Simone Arreghini.

Figure 1
Figure 1. Figure 1: Front and back view of the original Genny Zero platform. wheelchairs, which typically require manual propulsion or joystick-based con￾trol, Genny Zero is operated mainly through body-weight shifting. This allows the user to control the longitudinal motion of the platform in an intuitive and hands-free manner. The operating principle of Genny Zero is analogous to that of other self￾balancing systems [9] lik… view at source ↗
Figure 2
Figure 2. Figure 2: 3D, front and side views of the modified Genny Zero platform equipped with the LiDAR and RGB-D sensors, along with the laptop used for sensor acquisition and the navigation controller. The sensors’ FOV are highlighted in red for the ZED 2i and blue for the Robosense Airy LiDAR. Genny Zero weighs 70 kg and supports a payload of up to 120 kg. It includes redundant safety components, such as dual-winding moto… view at source ↗
Figure 3
Figure 3. Figure 3: Representative snapshots of the gesture-based people-following behavior. The top-left image shows the follow me gesture used to trigger people following, while the bottom-right image shows the stop gesture used to stop the Genny Zero platform. The accompanying video showing the complete execution of the presented experiments is available at: https://youtu.be/LVAix_Qx7bM. 3.1 People-following In the first e… view at source ↗
Figure 4
Figure 4. Figure 4: Representative snapshots of the people-following behavior in a constrained en￾vironment, including door crossing and corridor navigation. Genny trajectory Tracked user trajectory Genny Start Genny End User Start User End 2 m [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Top-down view of the trajectories recorded during the constrained indoor nav￾igation experiment. The red trajectory represents Genny Zero, while the green trajec￾tory represents the tracked user. performs the stop gesture, detected as both hands being close to their respective shoulders for the configured hold duration. In the second experiment, we extend the people-following behavior with ob￾stacle avoida… view at source ↗
Figure 6
Figure 6. Figure 6: Representative moment of the hailing behavior (left), along with the corre￾sponding LiDAR-based map used for obstacle avoidance (center) and the onboard RGB-D camera view (right). The top left frame shows an example of the hailing ges￾ture used to start the hailing behavior. 3.2 Hailing In the third experiment, we demonstrate hailing as a practical autonomous func￾tion that allows the user to call the whee… view at source ↗
read the original abstract

Wheelchair users call for assistive mobility systems that provide active support, adapt to dynamic environments, and are intuitive and user-friendly. However, powered wheelchairs typically still provide limited autonomy and lack effective integration with advanced perception and navigation capabilities, particularly in complex real-world environments. This paper presents a preliminary study toward autonomous powered wheelchairs for real-world assistive mobility. We introduce a proof-of-concept prototype that integrates autonomous perception, gesture-based interaction, and navigation on a commercially available self-balancing powered wheelchair. The proposed system builds upon Genny Zero, a commercial self-balancing wheelchair that enables hands-free and intuitive operation through body-weight shifting. To extend its capabilities toward autonomous operation, we integrate an RGB-D camera for human-aware perception and interaction, together with a LiDAR sensor for localization and navigation. We demonstrate the integrated system in two assistive applications: (i) hailing, allowing users to call the wheelchair from a distance; and (ii) people-following, where the wheelchair follows a person using leader-follower strategies, including a constrained indoor navigation example. The results highlight the potential of combining autonomous robotics with assistive mobility platforms, while also showing the feasibility of the proposed integration and identifying the main technical challenges that must be addressed before moving toward user-ready, accessible, and intelligent mobility solutions. A video demonstrating the experimental setup and results is available at: https://youtu.be/LVAix_Qx7bM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 7 minor

Summary. This paper presents a proof-of-concept integration of autonomous perception (LiDAR + RGB-D), gesture-based interaction, and navigation on the Genny Zero, a commercially available self-balancing powered wheelchair. The system is built on ROS 2 Nav2 with a custom driver that converts velocity commands into pitch and steering inputs for the self-balancing platform. Two assistive behaviors are demonstrated indoors: people-following (including constrained door-crossing and corridor navigation) and remote hailing. The paper is explicitly framed as a preliminary study and is transparent about its limitations: the wheelchair was unoccupied, motor torque was limited to 1 Nm, experiments were indoor-only, and no quantitative benchmarks are provided. The core technical contribution is the software-control interface and the integration pipeline rather than a novel algorithm or theoretical result.

Significance. The paper bridges a commercial self-balancing mobility platform with standard robotic navigation and perception modules, which is a practical engineering contribution of interest to the assistive robotics community. The inclusion of a video demonstrating the experiments and the honest discussion of integration challenges (especially the pitch-coupling issue) are commendable and add value. The work does not claim theoretical novelty; its significance lies in documenting the feasibility and limitations of the approach on a real platform. The gesture recognition module is simple but functional and is appropriately described as a first step.

major comments (3)
  1. Section 4, paragraph on navigation limitations: The paper acknowledges that the quasi-static planar approximation causes stop-and-go oscillations that interact with the self-balancing controller. This is the central load-bearing concern: the feasibility claim for the software-control interface rests on this approximation, yet the paper's own evidence shows instability symptoms even under the most conservative operating conditions (no payload, 1 Nm torque limit, 0.65 m/s). The paper should more explicitly characterize the severity and frequency of these oscillations—e.g., in how many of the conducted experiments did they occur, and did they cause any experiment to be aborted? This information is needed to assess whether the 'sufficient for proof-of-concept' claim is supported by the evidence.
  2. Section 3: The paper states that experiments are 'intended as preliminary demonstrations rather than as a quantitative benchmark,' but no quantitative results of any kind are provided—no success rates, no trajectory tracking errors, no timing data, no gesture recognition accuracy. Even for a proof-of-concept paper, minimal quantitative characterization (e.g., number of trials, success/failure counts, following-distance error) would substantially strengthen the feasibility claim and distinguish the demonstrated behaviors from single cherry-picked runs.
  3. Section 2.3, gesture recognition: The gesture recognition module is described as based on 'simple hands position rules,' but the specific thresholds, hold durations, and failure modes are not specified. Given that gesture recognition is the trigger for all demonstrated behaviors, at least a brief description of the detection logic and its robustness during the experiments would strengthen the system description.
minor comments (7)
  1. Section 2.1: The maximum velocity is stated as 20 km/h (approximately 5.55 m/s), hardware-limited to 10 km/h, and software-limited to 1.0 m/s. The Nav2 desired linear velocity is then stated as 0.65 m/s in Section 2.3. Clarify which velocity was actually used in experiments.
  2. Section 2.2: The sensor mount inclination (20 degrees) and the resulting ground-plane intersection distances are described, but a figure showing the actual fields of view overlaid on the platform (as partially done in Fig. 2) with the ground intersection points marked would improve clarity.
  3. Section 2.3: The command timeout of 0.25 s is mentioned as a safety mechanism. It would help to clarify whether this timeout was triggered during the experiments and whether it contributed to the stop-and-go oscillation issue.
  4. References [10, 16, 20] are citations to commercial product websites (DAAV, Omeo, Scewo) accessed in 2026. These may not be stable archival references; consider adding archived snapshots or finding published descriptions of these platforms.
  5. Section 3.1: The following distance of 1.25 m is mentioned in the text but the basis for this choice is not discussed. A brief justification would be useful.
  6. Fig. 5: The trajectory plot shows a '2 m' scale bar but the axes are unlabeled. Adding axis labels or a grid would improve interpretability. Also clarify whether the scale bar is 2 m (is it a scale bar or a dimension?).
  7. The abstract states 'A video demonstrating the experimental setup and results is available.' It would be appropriate to state how many trials were conducted and whether the video shows representative or cherry-picked runs.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee correctly identifies the paper as a proof-of-concept engineering contribution, and all three major comments request additional quantitative or technical detail that we agree would strengthen the manuscript. We address each comment below.

read point-by-point responses

  1. Referee: Section 4: Characterize severity and frequency of stop-and-go oscillations; state how many experiments were affected and whether any were aborted.

    Authors: The referee is correct that the oscillation issue is load-bearing for the feasibility claim, and we agree that the current description is insufficiently quantitative. We will revise Section 4 to include the following information. Across the experiments conducted (people-following in open space, people-following through doorway and corridor, and hailing), the stop-and-go oscillations occurred intermittently during the people-following experiments, particularly during the constrained doorway-crossing scenario where frequent replanning was triggered by jitter in the tracked user's position estimate. No experiment was aborted as a result of these oscillations; all demonstrated runs completed successfully. The oscillations were most pronounced during rapid direction changes and goal updates, and did not occur during the hailing experiment, where the navigation goal was static. We will add this characterization to the revised manuscript and note that the oscillations degraded motion smoothness but did not compromise safety or task completion under the conservative operating conditions used. revision: yes

  2. Referee: Section 3: No quantitative results at all; provide minimal metrics such as trial counts, success/failure rates, following-distance error, timing data, or gesture recognition accuracy.

    Authors: We agree that even a proof-of-concept paper benefits from minimal quantitative characterization, and we will add this to the revised manuscript. Specifically, we will report: (1) the number of trials conducted for each scenario (people-following in open space, people-following in constrained indoor navigation, and hailing), with success/failure counts; (2) following-distance error statistics from the people-following experiments, computed as the deviation between the achieved wheelchair-to-user distance and the configured 1.25 m setpoint; and (3) the trajectory data already shown in Figure 5 will be accompanied by quantitative tracking error metrics. We note that the current experimental setup was not designed as a systematic benchmark, so the quantitative data we can add is limited in scope. However, we agree it is sufficient and necessary to support the feasibility claim and to distinguish the results from cherry-picked runs. revision: yes

  3. Referee: Section 2.3: Specify gesture recognition thresholds, hold durations, and failure modes; describe robustness during experiments.

    Authors: We agree this information should be included and will add it to Section 2.3. Specifically, we will describe: (1) the detection logic for each of the three gestures (hailing: one hand above a height threshold relative to the head; follow me: one hand within a distance threshold of the corresponding shoulder; stop: both hands within the shoulder distance threshold); (2) the hold duration requirement (the gesture must be maintained for a configurable duration, which we will specify, before the behavior is triggered, to reduce false positives from transient poses); and (3) observed failure modes during the experiments, including cases where partial body occlusion or the user turning away from the camera caused tracking loss and gesture detection failure. We will also briefly comment on the robustness observed during the experiments, noting that gesture recognition was generally reliable when the user faced the camera and remained within the effective tracking range, but was susceptible to occlusion and pose ambiguity at close range or when the user was in motion. revision: yes

Circularity Check

0 steps flagged

No circularity: this is a systems-integration paper with no theoretical derivation chain to be circular.

full rationale

The paper presents a proof-of-concept integration of off-the-shelf components (Nav2, AMCL, Regulated Pure Pursuit, ZED SDK body tracking) on a commercial self-balancing wheelchair. It makes no first-principles theoretical claims, derives no equations whose outputs are defined by their inputs, and presents no fitted parameters as predictions. The self-citations [3,4,5,6] refer to prior HRI perception work by the authors but appear only in the future-work discussion (Section 4) as directions for richer capabilities—not as load-bearing premises for the integration claims demonstrated in Section 3. The quasi-static planar approximation (Section 2.3) is a modeling assumption whose limitations the paper openly acknowledges, not a circular reduction. The feasibility claim rests on empirical demonstrations (video, trajectory plots), not on a derivation that assumes its own conclusion. No circularity is present.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

The paper introduces no new theoretical entities, particles, forces, or mathematical constructs. All hardware and software components are commercial or open-source. The free parameters are engineering configuration values chosen for conservative proof-of-concept operation, not fitted to data. The axioms are domain assumptions about the adequacy of approximations for the demonstrated scenarios.

free parameters (8)
  • Following distance (people-following) = 1.25 m
    Chosen distance for the people-following behavior; not derived from any optimization or user study.
  • Hailing approach distance = 1.0 m
    Chosen final distance when approaching a hailing user; not derived.
  • Desired linear velocity (Nav2) = 0.65 m/s
    Conservative velocity limit set for proof-of-concept demonstrations.
  • Velocity smoother limits = 0.75 m/s, 0.95 rad/s
    Manually configured limits for the Nav2 velocity smoother.
  • Acceleration/deceleration limits = 1.20 m/s^2, -0.70 m/s^2
    Manually configured for conservative operation.
  • Inflation radius (costmap) = 0.7 m
    Standard Nav2 parameter chosen for obstacle avoidance; not tuned via systematic evaluation.
  • Command timeout = 0.25 s
    Safety timeout chosen for the stop-command mechanism.
  • Sensor mount inclination = 20 degrees
    Physically determined by mounting position on footrest; affects sensor coverage geometry.
axioms (4)
  • domain assumption The Genny Zero can be approximated as a planar differential-drive robot for low-speed navigation.
    Section 2.3: AMCL uses differential-drive motion model; pitch is removed from navigation representation. This is the foundational approximation enabling the entire autonomy stack.
  • domain assumption ZED 2i body tracking provides sufficiently accurate 3D joint positions for gesture recognition and people-following at the demonstrated ranges.
    Section 2.3 and 3.1: the system relies on pelvis position estimates for navigation goals and hand positions for gestures, without quantitative evaluation of tracking accuracy.
  • domain assumption A virtual 2D scan extracted from the 3D LiDAR point cloud is sufficient for navigation and obstacle avoidance in the demonstrated environments.
    Section 2.3: the 3D LiDAR data is reduced to a planar slice for Nav2, discarding 3D structure. The paper acknowledges this as a limitation in Section 4.
  • domain assumption Pre-built maps without people are reusable for localization during autonomous operation with people present as temporary obstacles.
    Section 2.3: 'mapping is performed in environments without people. The resulting map is then saved and reused during execution.' This assumes the static structure dominates localization.

pith-pipeline@v1.1.0-glm · 12464 in / 3138 out tokens · 267335 ms · 2026-07-08T07:34:45.010239+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Robotics and Au- tonomous Systems171, 104568 (2024)

    Abbate, G., Giusti, A., Schmuck, V., Celiktutan, O., Paolillo, A.: Self-supervised prediction of the intention to interact with a service robot. Robotics and Au- tonomous Systems171, 104568 (2024)

    work page 2024

  2. [2]

    Autonomous Robots47(8), 1255–1273 (2023)

    Antonucci, A., Bevilacqua, P., Leonardi, S., Paolopoli, L., Fontanelli, D.: Humans as path-finders for mobile robots using teach-by-showing navigation. Autonomous Robots47(8), 1255–1273 (2023)

    work page 2023

  3. [3]

    In: ACM/IEEE International Conference on Human-Robot Interaction

    Arreghini, S., Abbate, G., Giusti, A., Paolillo, A.: A long-range mutual gaze detec- tor for hri. In: ACM/IEEE International Conference on Human-Robot Interaction. pp. 870–874 (2024)

    work page 2024

  4. [4]

    In: IEEE International Conference on Robotics and Automation

    Arreghini, S., Abbate, G., Giusti, A., Paolillo, A.: Predicting the intention to inter- act with a service robot: the role of gaze cues. In: IEEE International Conference on Robotics and Automation. pp. 993–999 (2024) 14 S. Arreghini et al

    work page 2024

  5. [5]

    In: IEEE/RSJ International Conference on Intelligent Robots and Systems

    Arreghini, S., Abbate, G., Giusti, A., Paolillo, A.: A service robot in the wild: Analysis of users intentions, robot behaviors, and their impact on the interaction. In: IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2536–2541 (2024)

    work page 2024

  6. [6]

    In: IEEE International Conference on Advanced Robotics and its Social Impacts

    Arreghini, S., Carlotti, N., Nava, M., Paolillo, A., Giusti, A.: Sixth-sense: Self- supervised learning of spatial awareness of humans from a planar lidar. In: IEEE International Conference on Advanced Robotics and its Social Impacts. pp. — (2026)

    work page 2026

  7. [7]

    International Journal of Social Robotics8(3), 393–408 (2016)

    Bonaccorsi, M., Fiorini, L., Cavallo, F., Saffiotti, A., Dario, P.: A cloud robotics so- lution to improve social assistive robots for active and healthy aging. International Journal of Social Robotics8(3), 393–408 (2016)

    work page 2016

  8. [8]

    European Journal of Operational Research277(2), 396–411 (2019)

    Boysen, N., De Koster, R., Weidinger, F.: Warehousing in the e-commerce era: A survey. European Journal of Operational Research277(2), 396–411 (2019)

    work page 2019

  9. [9]

    In: IEEE International Conference on Mechatronics

    Brentari, M., Zambotti, A., Zaccarian, L., Bosetti, P., Biral, F.: Position and speed control of a low-cost two-wheeled, self-balancing inverted pendulum vehicle. In: IEEE International Conference on Mechatronics. pp. 347–352 (2015)

    work page 2015

  10. [10]

    DAAV SA: Autonomous mobility solutions.https://www.daav.ch/, accessed: 2026-04

    work page 2026

  11. [11]

    Urban Planning2(2), 13–25 (2017)

    Jittrapirom, P., Caiati, V., Feneri, A.M., Ebrahimigharehbaghi, S., Alonso- González, M.J., Narayan, J.: Mobility as a service: A critical review of definitions, assessments of schemes, and key challenges. Urban Planning2(2), 13–25 (2017)

    work page 2017

  12. [12]

    IEEE Access12, 164248–164261 (2024)

    Leong, P.Y., Ahmad, N.S.: Lidar-based obstacle avoidance with autonomous vehi- cles: A comprehensive review. IEEE Access12, 164248–164261 (2024)

    work page 2024

  13. [13]

    IEEE Robotics and Automation Letters (2025)

    Liao, Y., Xu, X., Bai, R., Yang, Y., Cao, M., Yuan, S., Xie, L.: Following is all you need: Robot crowd navigation using people as planners. IEEE Robotics and Automation Letters (2025)

    work page 2025

  14. [14]

    Autonomous Robots47(6), 685–694 (2023)

    Macenski, S., Singh, S., Martín, F., Ginés, J.: Regulated pure pursuit for robot path tracking. Autonomous Robots47(6), 685–694 (2023)

    work page 2023

  15. [15]

    ACM Transactions on Human-Robot Interaction12(3), 1–39 (2023)

    Mavrogiannis, C., Baldini, F., Wang, A., Zhao, D., Trautman, P., Steinfeld, A., Oh, J.: Core challenges of social robot navigation: A survey. ACM Transactions on Human-Robot Interaction12(3), 1–39 (2023)

    work page 2023

  16. [16]

    Omeo Technology: About us.https://omeotechnology.com/about-us/, accessed: 2026-04

    work page 2026

  17. [17]

    World Health Organi- zation (2022)

    Organization, W.H.: Global report on assistive technology. World Health Organi- zation (2022)

    work page 2022

  18. [18]

    arXiv preprint arXiv:2511.05158 (2025)

    Salimpour, S., Catalano, I., Westerlund, T., Falahi, M., Queralta, J.P.: Follow- me in micro-mobility with end-to-end imitation learning. arXiv preprint arXiv:2511.05158 (2025)

    work page arXiv 2025

  19. [19]

    International Journal of Social Robotics 14(2), 441–462 (2022)

    Salvini, P., Paez-Granados, D., Billard, A.: Safety concerns emerging from robots navigating in crowded pedestrian areas. International Journal of Social Robotics 14(2), 441–462 (2022)

    work page 2022

  20. [20]

    Scewo AG: About scewo.https://www.scewo.com/en/about/, accessed: 2026-04

    work page 2026

  21. [21]

    Communications of the ACM56(1), 116–124 (2013)

    Shotton, J., Sharp, T., Kipman, A., Fitzgibbon, A., Finocchio, M., Blake, A., Cook, M., Moore, R.: Real-time human pose recognition in parts from single depth images. Communications of the ACM56(1), 116–124 (2013)

    work page 2013