Researchers from Germany's DFKI have demonstrated prototype electric wheelchairs that can navigate obstacle-filled rooms using natural-language voice commands, presenting their findings at the CSUN Assistive Technology Conference in Anaheim, California, earlier this month.
The project, called Reliable and Explainable Swarm Intelligence for People With Reduced Mobility (REXASI-PRO), was co-led by senior researcher Christian Mandel and colleague Serge Autexier at DFKI's Bremen facility. The wheelchairs operate in two modes: a semiautonomous mode using a joystick with AI assistance, and a fully autonomous mode where users issue spoken destinations — "Please drive me to the coffee machine" — and the chair handles the rest.
How the System Actually Works
Each of the two prototype wheelchairs used in experiments was equipped with two lidar sensors, a 3D camera, odometers, a user interface, and an embedded computer. Autonomous navigation runs on the open-source ROS2 Nav2 system, which processes natural-language input, builds simultaneous localization and mapping (SLAM) maps of the environment, and deploys local obstacle-avoidance controllers to adjust the chair's path in real time.
The researchers also tested a supplementary safety layer that fused sensor data from both the wheelchair and external room-based sensors, including drone-mounted color and depth cameras. In practice, a user presses a key on the human-machine interface, speaks a destination command, confirms or rejects the interpreted instruction, and the chair then navigates while continuously scanning for obstacles.
"Never underestimate what wheelchair users can do without it," — Christian Mandel, DFKI senior researcher.
That reflection from Mandel — drawn from early-career experiments watching severely disabled users thread narrow passages with precision — captures the central tension in the field: the technology must genuinely help, not merely replace capabilities users already possess.
The Real Barriers Aren't Technical Alone
Outside researchers point to three persistent obstacles beyond engineering. Pooja Viswanathan, CEO and founder of Toronto-based Braze Mobility, identifies cost as the first. "Funding systems are often not designed to support advanced add-on intelligence unless there is very clear evidence of value and safety," she says. Reliability in messy, variable real-world conditions is the second. The third is human factors: users differ widely in cognitive, motor, sensory, and environmental needs, making universal solutions difficult to design.
Braze's own product philosophy reflects this tension directly. Rather than pursuing full autonomy, the company makes blind-spot sensor add-ons that attach to any existing electric wheelchair and deliver multimodal alerts to users — supporting human decision-making rather than supplanting it.
Louise Devinge, a biomedical research engineer at IRISA (Research Institute of Computer Science and Random Systems) in Rennes, France, raises a systems-level concern. "The more sensing, computation, and autonomy you add," she says, "the harder it becomes to ensure robust performance across the full range of real-world environments that wheelchair users encounter." Greater capability, in other words, introduces greater complexity — and complexity is the enemy of the reliability that assistive technology demands.
Where the Research Sits on the Spectrum
Viswanathan characterizes the REXASI-PRO system as sitting at the ambitious end of the smart wheelchair spectrum — beyond what today's commercial products can deliver, but valuable precisely because of that. "Its strengths appear to lie in intelligent navigation, advanced sensing, and the broader effort to build a wheelchair that can interpret and respond to complex environments in a more autonomous way," she says.
She also singles out the project's explicit focus on trustworthy and explainable AI as appropriate for a safety-critical mobility context. When a device physically transports a person, the consequences of a failure are immediate and potentially serious — making the 'why' of an AI decision as important as the decision itself.
Mandel projects that smart wheelchairs ready for mainstream consumers are roughly 10 years away. That timeline implicitly acknowledges that the gap between laboratory demonstration and daily-life reliability remains significant — a gap that researchers, clinicians, regulators, and funders will all need to close together.
The REXASI-PRO work was supported in part by the IEEE Foundation and a Jon C. Taenzer fellowship grant. All performance claims reflect the research team's own experimental findings and have not been independently verified.
What This Means
For wheelchair users, clinicians, and assistive technology developers, this research marks a credible step toward AI-assisted mobility — but the field's immediate priority is building systems reliable and affordable enough for daily life, not pursuing full autonomy for its own sake.