TL;DR
This paper introduces a safety-first control system for electric wheelchairs that uses Robust Adaptive Model Predictive Control (RAMPC) to prevent dangerous speeds and accelerations. By combining online parameter estimation with a "safety filter" for joystick inputs, the controller ensures the wheelchair remains stable and within safety limits even when navigating ramps or carrying users of different weights.
What Problem Does This Paper Address?
Electric wheelchairs are essential for mobility, but they are often controlled by raw joystick inputs that don't account for the laws of physics or safety boundaries. A sudden push on the joystick can cause jerky, high-acceleration movements that might tip the chair or injure the user.
Furthermore, wheelchairs face two major sources of uncertainty:
- External Disturbances: Navigating inclined ramps or uneven terrain adds unpredictable forces.
- Parameter Variation: A wheelchair's dynamics change significantly depending on the weight of the user. A controller tuned for a 60kg user might perform poorly or unsafely for a 90kg user.
Existing controllers often ignore these dynamics or fail to guarantee that safety constraints (like maximum acceleration) will be met in real-world, "noisy" conditions.
Methods
The authors propose a multi-layered control architecture:
- Linear Parameter Varying (LPV) Model: They model the wheelchair's actuators and mechanics at the dynamic level (voltage to velocity/acceleration).
- Set-Membership Parameter Estimation: Instead of guessing a single value for system parameters (like mass or friction), the system maintains a "set" of possible values. It updates this set online as it gathers data, allowing the chair to "learn" the current user's characteristics.
- Admissible Reference Generation: If a user moves the joystick in a way that would violate safety limits (an "inadmissible reference"), the controller automatically calculates the closest possible safe speed and steers toward that instead.
- Tube-based Robust MPC: The controller uses a "tube" approach to handle disturbances. It plans a central path but ensures that even with the worst-case disturbances, the wheelchair stays within a safe "tube" of operation.
Datasets
The researchers did not use a public dataset. Instead, they performed system identification on a physical electric wheelchair equipped with Brushless DC motors and absolute encoders.
- They collected voltage and velocity data from a 80kg participant driving with a joystick.
- This data was used to identify motor parameters ($l, r, k_e, k_t$) and nominal mass/damping matrices.
- The final evaluation was conducted on a high-fidelity wheelchair model that simulated three different user weights (60kg to 85kg) to test the adaptive capabilities.
Key Insights and Ideas
- Safety as a Filter: The core idea is that the controller acts as a "guardian" between the user and the motors. It respects the user's intent but prioritizes physical safety constraints.
- Handling Infeasibility: Most MPC controllers simply fail if a target is impossible. This system uses an optimization-based method to find the "closest admissible" target, ensuring the chair remains responsive even when the user asks for something unsafe.
- Online Adaptation for Shared Mobility: The authors highlight that this is particularly useful for "rentable" or shared wheelchairs in hospitals, where the system must automatically adjust to a new user's weight without manual recalibration.
Key Results
The RAMPC was compared against a standard Proportional-Integral (PI) controller in two tasks:
- Task A (Ramps): When climbing and descending 5-degree ramps, the RAMPC successfully kept acceleration within the 2 m/s² safety limit. In contrast, the PI controller caused the wheelchair to exceed acceleration tolerances and introduced velocity oscillations.
- Task B (Sinusoidal Tracking): When given a target speed that fluctuated in and out of the "safe zone," the RAMPC smoothly tracked the reference when safe and "clipped" the movement at the safety boundary when unsafe.
- Precision: The RAMPC achieved a maximum speed tracking error of approximately 0.048 m/s, significantly better than the PI controller's 0.158 m/s.
Limitations and Future Work
- Computational Load: While the authors used techniques to reduce the online computation load, Model Predictive Control is inherently more "expensive" to run on low-power embedded wheelchair hardware than simple PI loops.
- Static Obstacles: This paper focuses on speed and acceleration safety (internal state). It does not explicitly integrate external obstacle avoidance (like LiDAR or camera-based braking), though the framework could theoretically be extended to include distance-to-obstacle constraints.
- Real-world Testing: While the model was high-fidelity and based on real hardware parameters, the final "safety" tests were performed in simulation. Future work would involve testing these edge-case safety violations with human subjects in a controlled environment.
Summary updated