CoNav Chair: Shared-Control Smart Wheelchair

• CoNav Chair is a ROS-based shared-control smart wheelchair that balances user input with autonomous planning for indoor mobility.
• It integrates multi-sensor perception, graph-based mapping, LiDAR–IMU odometry, and a nonlinear MPC planner to optimize navigation safety and efficiency.
• Experimental results show that shared control reduces collision risks and control burden while enhancing user satisfaction through blended navigation modalities.

Searching arXiv for the CoNav Chair paper and closely related work to ground the article in the current literature. I’m checking the arXiv entries most directly related to CoNav Chair, including the original system paper and the later evaluation paper, to ensure the terminology and claims align with the literature. CoNav Chair is a ROS-based smart wheelchair for shared-control navigation in the built environment, developed on a commercial powered wheelchair platform to bridge the gap between pure manual driving and full autonomy. Its defining characteristic is the integration of multi-sensor perception, graph-based mapping, LiDAR–IMU odometry, obstacle avoidance, and a shared-control nonlinear Model Predictive Control (MPC) local planner that blends user intent with an autonomous global path rather than assigning exclusive authority to either the human or the machine. In the authors’ framing, the central problem is not only navigation performance, but the joint preservation of efficiency, comfort, retained agency, and user trust, all of which are treated as prerequisites for the acceptance of assistive mobility technologies in indoor built environments (Xu et al., 16 Jan 2025). A later paper extends the system description with a preliminary usability evaluation comparing manual, autonomous, and shared modes in a university corridor with 21 healthy, unimpaired participants (Xu et al., 15 Jul 2025).

1. Design rationale and research context

CoNav Chair was introduced in response to a persistent bifurcation in powered wheelchair research and practice. Commercial powered wheelchairs commonly expose manual joystick control, while robotic wheelchair research has often emphasized either full autonomy or enhanced manual control. The CoNav Chair papers argue that both poles have characteristic failure modes indoors: full autonomy can underperform in complex layouts with sharp turns, centrally positioned obstacles, sensor field-of-view limitations, and “freezing robot” behaviors, whereas user-only control increases cognitive and fine-motor demand and can become fatiguing and unreliable over time. The papers also emphasize that many commercial powered wheelchairs lack basic perception sensors such as LiDAR and RGB-D, forcing users to compensate for limited situational awareness (Xu et al., 16 Jan 2025).

The system therefore occupies an intermediate design space. Perception, mapping, localization, and safe execution are delegated to the autonomy stack, while the user retains the ability to influence path and motion decisions. This positioning is fundamental to the project’s interpretation of shared control: CoNav Chair is not presented as an autonomy-first wheelchair with occasional override, nor as a manually operated wheelchair with added collision filtering. Rather, it is a platform in which user intent and autonomous planning are continuously reconciled inside the local planner.

Within the prior literature identified by the authors, systems such as NavChair, Vulcan 2.0, and Cyberwork Wheelchair exemplify autonomy-oriented trajectories, while Wheelsley exemplifies enhanced manual control through eye tracking. CoNav Chair is presented as differing from these earlier systems by combining SLAM-capable mapping, autonomous navigation, and shared control on a commercial powered wheelchair, using a sensing stack that includes both 3D LiDAR and RGB-D/stereo perception (Xu et al., 16 Jan 2025).

2. Hardware platform and sensing configuration

The hardware base is a Quantum Q6 Edge 2.0 powered wheelchair with an R-Net base and enhanced display/alternative control interface. The differential-drive powered wheelchair motors and the integrated electromagnetic brakes are retained. A custom motor controller board outputs two analog voltages in the 4.8–7.2 V range to the wheelchair’s alternative control interface, corresponding to forward and rotational velocity channels; on-board digital potentiometers convert ROS commands into these voltages, and the alternative control interface converts them into CAN messages for the R-Net base. A microcontroller on the custom motor controller board communicates with the ROS computer through rosserial, serving as the bridge to the alternative control interface. The computing unit is a laptop PC running Ubuntu 20.04 and ROS Noetic, connected to the sensors and motor controller. The wheelchair-integrated batteries remain in their standard powered-wheelchair configuration and are not modified in the paper (Xu et al., 16 Jan 2025).

The perception package comprises a Velodyne VLP-16 3D LiDAR, a ZED stereo camera, a Witmotion IMU, and two retrofitted motor-shaft wheel encoders. The LiDAR is mounted via a bracket to T-slotted aluminum extrusions attached to the armrest and seat rail. The ZED stereo camera is mounted on a pivot joint attached to the extrusions. The wheel-encoder retrofit required the motor shafts to be drilled and threaded to accommodate encoder mounts, while preserving the existing mounting points and electromagnetic brakes. Cables are routed along the extrusions and secured with zip ties to avoid interference with leg movement.

The communication architecture is correspondingly layered. Serial communication via rosserial links the ROS laptop and the motor controller board. CAN communication remains internal to the R-Net base. Sensor connectivity to the laptop is via standard interfaces such as USB, although the exact cabling is not enumerated. The papers do not report exact sensor sampling rates or calibration procedures, but they state that RTAB-Map uses RGB-D/stereo data in real time and that Faster-LIO fuses LiDAR and IMU at real-time rates appropriate for indoor operation (Xu et al., 16 Jan 2025).

The later evaluation paper adds interface details absent from the earlier design description: manual driving and on-demand intervention in shared mode are implemented with an Xbox joystick controller, while a touch-based user interface is used to set goals in autonomous and shared modes (Xu et al., 15 Jul 2025).

3. ROS software stack and shared-control navigation formulation

The software stack is built on Ubuntu 20.04 and ROS Noetic. RTAB-Map provides graph-based SLAM, producing a 3D map and a 2D projection published through a map server on the topic /map. Faster-LIO supplies tightly coupled LiDAR–IMU odometry on /odom. A motor-control ROS node sends command velocity signals to the microcontroller through rosserial; the board maps these commands to analog voltages that are then converted by the wheelchair interface into CAN messages for the base. The papers describe a standard ROS velocity topic, such as /cmd_vel with a message type typically geometry_msgs/Twist, although the topic name is not explicitly fixed in the first paper. TF frames are not enumerated, but map, odom, and base_link are described as typical for such a setup. The system does not use standard move_base, costmap2d, DWA, or TEB; navigation instead depends on a custom MPC-based local planner (Xu et al., 16 Jan 2025).

The local planner is a nonlinear MPC formulation that tracks a blended reference trajectory. The optimization is stated as

$$\mathbf{u}^{*} = \arg\min_{\mathbf{u}\in\mathcal{U}} \mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u})$$

subject to

$$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$

The cost is decomposed into state-tracking and control-effort terms:

$$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$

$$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$

$$\mathcal{J}_u(\mathbf{u}) = \mathbf{u}^{T}\mathbf{Q}_{u}\mathbf{u}.$$

The shared reference is

$$\hat{\mathbf{s}} = \theta(k)\,\mathbf{s}^{\text{user}} + \big(1 - \theta(k)\big)\,\mathbf{s}^{\text{global}},$$

where $\mathbf{s}^{\text{user}}$ is a predicted state sequence derived from user control, $\mathbf{s}^{\text{global}}$ is the global-plan reference, and $\theta(k)\in[0,1]$ depends on $k$, the number of user control signals over a time window. The first paper gives the exponential form

$$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$0

while also stating that $$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$1 lies in $$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$2 and increases with user input frequency. This suggests a sign inconsistency in the printed mapping. The later paper explicitly characterizes the printed expression as a sign typo and proposes a bounded monotone implementation such as $$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$3 with clipping to $$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$4 (Xu et al., 15 Jul 2025).

The wheelchair is modeled as a differential-drive platform, with standard planar kinematics

$$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$5

Collision avoidance is expressed through the constraint $$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$6, while control feasibility is enforced through $$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$7; exact velocity and rotational bounds are not reported. Smoothness and comfort are encouraged through the control penalty $$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$8, and low-level execution is stabilized by a closed-loop PID motor-control layer (Xu et al., 16 Jan 2025).

A conceptually important point in the later paper is that CoNav Chair does not directly blend user and robot commands. It blends reference trajectories and then solves an MPC problem over the blended reference. This distinction matters because it situates shared control within a dynamics- and constraint-aware optimizer rather than as an algebraic interpolation of joystick and autonomous commands (Xu et al., 15 Jul 2025).

4. Interaction model, operational modes, and safety architecture

The human–robot interface preserves a strong notion of user agency. In the original design description, the system retains the powered wheelchair’s user controls via the R-Net enhanced display and joystick interface, and the user’s control-signal sequence is incorporated into the shared-control planner as $$s_{t+1} = f(s_t, u_t), \quad \mathbf{s}\in\mathcal{S}_{\text{free}}, \quad \mathbf{u}\in\mathcal{U}.$$9 (Xu et al., 16 Jan 2025). In the later evaluation paper, the interaction design is formalized into three operational modes. In manual mode, the user drives exclusively with the Xbox joystick. In autonomous mode, the user selects a goal on the touch-based user interface and the wheelchair navigates without user command inputs. In shared mode, the user selects a goal and may intervene at any time with the joystick; the weight $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$0 scales user influence in the local planner as a function of input frequency (Xu et al., 15 Jul 2025).

This operational logic clarifies a common misconception about the system. CoNav Chair is not a fully autonomous wheelchair with a fallback manual override. Its shared mode is the primary conceptual contribution, and its authority-sharing mechanism is continuous rather than mode-switching at the level of single commands. The autonomy stack handles situational awareness, mapping, and safe trajectory generation; the human remains able to redirect behavior in ambiguous or preference-sensitive situations.

Safety is distributed across several layers rather than concentrated in a single formal certificate. The electromagnetic brakes of the base wheelchair are retained. The low-level PID controller improves command tracking and helps prevent oscillations. Collision avoidance is enforced through MPC feasibility, using the current map and obstacle representation maintained by RTAB-Map and the localization supplied by Faster-LIO. At the same time, the papers do not report collision watchdogs, recovery behaviors, maximum speed limits, braking distances, clearance margins, e-stop implementation, or standards compliance. The later paper explicitly notes that formal safety guarantees such as control barrier certificates are not claimed and points to a socially aware MPC-DCBF variant as future work (Xu et al., 15 Jul 2025).

The system is explicitly tailored to constrained indoor environments. Narrow corridors, clutter, and tight maneuvers are treated as core operating conditions rather than edge cases. Doorways, ramps, and elevators, however, are not explicitly tested in the reported studies.

5. Experimental evidence and measured performance

The first CoNav Chair paper reports experiments in a typical indoor social environment: a first-floor corridor in an academic building with four corridor scenes and obstacles distributed along the route. Three navigation modes were evaluated—fully manual control, fully autonomous navigation, and shared-control navigation—using the metrics of time to completion, success rate, collision rate, user satisfaction, trajectory length, and cumulative absolute angle difference as a smoothness measure. Participant demographics and detailed protocols are not reported. The reported outcome is qualitative but unambiguous: shared control outperformed manual and autonomous modes across all reported metrics, with shorter completion times, shorter trajectories, higher success rates, fewer collisions, and smoother paths. The authors interpret manual control as susceptible to collisions and lower success because it requires frequent adjustments near obstacles, while full autonomy can avoid obstacles yet become inefficient or freeze in complex layouts; shared control is presented as balancing both regimes by allowing users to adjust decisions in difficult situations while leaving perception and safety to the autonomy stack (Xu et al., 16 Jan 2025).

The later evaluation paper provides a more detailed preliminary usability study with $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$1 healthy, unimpaired adults aged 18–45, comprising 10 male and 11 female participants. Most had no prior wheelchair experience (95.2%), and 81.0% had limited secondhand exposure. The site was a university-building corridor of length approximately 25 m, with eight whiteboards arranged alternately on both sides to form a zig-zag and two inflatable manikins used as static pedestrian surrogates. Each participant completed two repetitions per mode, with training in a separate area before each mode. A researcher shadowed each run to log collisions and behaviors, and after each mode participants completed a 9-item, 5-point Likert questionnaire covering efficiency, safety, and ease of use (Xu et al., 15 Jul 2025).

The objective results are numerically specific. Completion time $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$2 was $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$3 s for manual mode, $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$4 s for autonomous mode, and $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$5 s for shared mode. Path length $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$6 was $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$7 m, $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$8 m, and $$\mathcal{J}(\mathbf{s}, \hat{\mathbf{s}}, \mathbf{u}) = \mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) + \mathcal{J}_u(\mathbf{u}),$$9 m, respectively, with a nonsignificant mode effect ($$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$0, $$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$1). Smoothness, measured by cumulative heading change

$$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$2

was $$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$3 rad for manual mode, $$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$4 rad for autonomous mode, and $$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$5 rad for shared mode, again without a significant mode effect ($$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$6, $$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$7). Control percentage—the proportion of the run with active joystick input—fell from $$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$8 in manual mode to $$\mathcal{J}_s(\mathbf{s}, \hat{\mathbf{s}}) = (\mathbf{s} - \hat{\mathbf{s}})^{T}\mathbf{Q}_{s}(\mathbf{s} - \hat{\mathbf{s}}),$$9 in shared mode, a large reduction in control burden ($$\mathcal{J}_u(\mathbf{u}) = \mathbf{u}^{T}\mathbf{Q}_{u}\mathbf{u}.$$0, $$\mathcal{J}_u(\mathbf{u}) = \mathbf{u}^{T}\mathbf{Q}_{u}\mathbf{u}.$$1). Collision totals across all trials were 27 for manual mode, 158 for autonomous mode, and 18 for shared mode, with a significant mode effect by Kruskal–Wallis test ($$\mathcal{J}_u(\mathbf{u}) = \mathbf{u}^{T}\mathbf{Q}_{u}\mathbf{u}.$$2, $$\mathcal{J}_u(\mathbf{u}) = \mathbf{u}^{T}\mathbf{Q}_{u}\mathbf{u}.$$3); pairwise comparisons showed shared $$\mathcal{J}_u(\mathbf{u}) = \mathbf{u}^{T}\mathbf{Q}_{u}\mathbf{u}.$$4 autonomous and manual $$\mathcal{J}_u(\mathbf{u}) = \mathbf{u}^{T}\mathbf{Q}_{u}\mathbf{u}.$$5 autonomous, while shared versus manual was not significant (Xu et al., 15 Jul 2025).

Subjective findings are consistent with the collision data. Shared control received median ratings of 4/5 on all nine questionnaire items. Significant safety-related differences favored shared and manual modes over autonomous mode on handling unexpected obstacles or changes, helping avoid hazards, and feeling safe in the mode. Fifteen of the 21 participants preferred shared control, six preferred manual control, and none preferred full autonomy. The reported limitation most frequently noted by participants was input-to-motion latency producing occasional overcorrections, attributed primarily to hardware and control-pipeline delays (Xu et al., 15 Jul 2025).

Taken together, the two papers support a stable empirical picture: shared control is presented as preserving the efficiency associated with autonomy while substantially reducing collision burden and user control effort.

6. Limitations, interpretation, and relation to adjacent work

Several boundaries of the current evidence are explicit. The first paper does not report participant demographics, detailed experimental protocols, ablations over $$\mathcal{J}_u(\mathbf{u}) = \mathbf{u}^{T}\mathbf{Q}_{u}\mathbf{u}.$$6, sensor-configuration studies, costmap parameters, planner timeouts, recovery behaviors, sampling rates, or calibration procedures. The later paper improves methodological specificity, but its evaluation remains preliminary because it is limited to healthy, unimpaired participants rather than powered-wheelchair users who rely on assistive mobility in daily life (Xu et al., 16 Jan 2025). The later paper therefore frames the reported study as groundwork for subsequent usability testing with people with disabilities who rely on powered wheelchairs (Xu et al., 15 Jul 2025).

A second limitation concerns formal safety claims. The system empirically reduces collisions and retains electromagnetic brakes and PID stabilization, but the papers do not report watchdogs, standards compliance, or theorem-level safety guarantees. This is significant because assistive robotics often demands stronger evidence than laboratory obstacle counts alone. A plausible implication is that CoNav Chair currently occupies the space of practically engineered shared autonomy rather than formally certified clinical assistive technology.

The most salient future directions stated or strongly implied in the papers are reduced input-to-actuation latency, richer intent inference, personalization of user models, adaptation of the authority-weighting scheme, outdoor operation, robustness to sensor field-of-view limitations, and socially aware navigation in dynamic crowds. The later paper specifically identifies an SS-MPC-DCBF algorithm as a target for future work on formal safety and social compliance (Xu et al., 15 Jul 2025).

The system’s name also invites a terminological clarification. CoNav Chair is distinct from the human-centered collaborative navigation benchmark “CoNav” (Li et al., 2024) and from the embodied vision-language-navigation framework “MA-CoNav” (Luo et al., 3 Mar 2026). Those works address intention-aware collaborative navigation in simulation and multi-agent VLN, respectively, whereas CoNav Chair concerns shared-control powered wheelchair navigation in indoor built environments. The overlap is lexical rather than methodological.

In summary, CoNav Chair defines a particular synthesis within smart-wheelchair research: a commercial powered wheelchair base, a ROS-centered perception-and-control stack, RTAB-Map and Faster-LIO for mapping and odometry, and a nonlinear shared-control MPC that blends user and global references. Its contribution is not only architectural. It also advances a specific normative claim about assistive mobility: that safe indoor navigation must preserve user authority rather than replace it, and that trust emerges from this balance of agency and robotic assistance rather than from autonomy alone (Xu et al., 16 Jan 2025).