Abstract
We introduce reachability analysis for the formal examination of robots. We propose a novel identification method, which preserves reachset conformance of linear systems. We additionally propose a simultaneous identification and control synthesis scheme to obtain optimal controllers with formal guarantees. In a case study, we examine the effectiveness of using reachability analysis to synthesize a state-feedback controller, a velocity observer, and an output feedback controller.
Introduction
Bridging the "reality gap" remains a significant challenge in robotics. While formal methods allow researchers to mathematically prove that a robot will avoid collisions or maintain specific error bounds, these guarantees are only valid if the underlying model accurately reflects the physical system. Real-world robots are subject to unmodeled dynamics such as joint flexibility, friction, sensor noise, and communication delays.
If a controller is designed for an idealized model, it may fail in practice. Conversely, overly conservative models result in sluggish performance. This work addresses this by unifying model identification and controller synthesis within a reachability framework to ensure "reachset conformance"—guaranteeing that a model's predicted reachable sets contain all real-world behaviors.
Methodology
The proposed framework leverages reachability analysis to bound all possible system behaviors using zonotopes—efficient geometric polytopes that facilitate fast Minkowski sums and linear transformations.
Reachset-Conformant Identification
Unlike standard regression techniques that minimize mean squared error, reachset-conformant identification ensures that the "reachable set" (the envelope of all possible states) encloses every measured data point. This ensures the model acts as a reliable "container" for physical reality.
Unified Optimization and Iterative Synthesis
The core of the approach is a unified optimization problem that simultaneously determines model parameters and controller gains. This allows for a trade-off between model tightness and controller aggressiveness. The process is iterative: initial data is used for synthesis, the new controller is deployed to collect more data (potentially exciting new dynamics), and the model is refined.
The authors proved that checking conformance can be reduced to a set of linear inequalities, making the optimization computationally efficient.
Experimental Results
The framework was validated using a Schunk LWA-4P 6-DOF robot manipulator. Over 33 minutes of high-frequency sensor data, comprising nearly 500,000 test cases, were used to verify reachset conformance.
Tighter Performance Guarantees
The unified approach significantly outperformed traditional sequential methods (identifying a model then designing a controller). For instance, on Axis 1, the unified method achieved a tracking error bound of 0.625 rad, whereas the separate approach was far more conservative.
As shown in the figure above, the reachable sets (blue bubbles) shrink over successive iterations $k$, providing increasingly precise formal guarantees while ensuring that the actual robot behavior remains enclosed within the predicted bounds. Furthermore, the method successfully maintained torque limits, a constraint often ignored in standard control designs like Linear-Quadratic Gaussian (LQG) control.
Conclusion and Future Work
This research demonstrates that unifying identification and control synthesis through reachability analysis provides robust, formal guarantees for physical robot systems. While currently focused on linear or feedback-linearized systems, future work aims to extend this framework to fully nonlinear models and more complex set representations like Taylor models. Additionally, the development of online conformance monitors could provide real-time safety triggers when unexpected dynamics are encountered.
Summary updated