Abstract
This paper presents a framework to imbue fast, neural-network-based Differentiable Predictive Control (DPC) with the rigorous safety guarantees characteristic of traditional Model Predictive Control (MPC). By leveraging a novel "sampled-data" Control Barrier Function (SD-ZCBFII), the authors develop a safety filter that intervenes only when the system approaches the boundaries of its safe operating region. This approach provides a computationally efficient alternative for high-speed applications where real-time optimization is impractical, maintaining safety even under external disturbances and sampling effects.
Introduction
Autonomous systems and robotics often face a fundamental trade-off between control frequency and safety. Model Predictive Control (MPC) is the industry standard for safety, as it solves an optimization problem at every time step to satisfy constraints. However, its high computational cost often precludes its use in high-speed applications. Conversely, learning-based controllers like DPC are extremely fast once trained but lack deterministic safety guarantees, often acting as "black boxes" in edge cases. This research bridges this gap by integrating Control Barrier Functions (CBFs) into a differentiable control pipeline.
Methodology: Differentiable Closed-Loop Control
The proposed architecture integrates a neural control policy within a differentiable closed-loop system. The framework utilizes forward propagation for system simulation and backward propagation for feedback-based training of the neural network.
SD-ZCBFII and Safety Filtering
The core theoretical innovation is the Sampled-Data Zeroing Control Barrier Function II (SD-ZCBFII). Traditional barrier functions can be overly conservative, restricting performance even when the system is far from danger. SD-ZCBFII addresses this by defining an "annulus" or buffer zone near the safety boundary. Safety constraints are strictly enforced only when the system enters this zone.
The system operates via an event-triggered safety filter:
- Normal Mode: The system runs the fast DPC neural network policy.
- Safety Mode: If the state enters the buffer zone, a Quadratic Program (QP) filter is triggered. This filter minimally adjusts the neural network's command to ensure the system remains within the safe set.
Training and Implementation
During offline training, the DPC loss function is augmented with barrier penalties. This "encourages" the neural network to learn a policy that naturally stays away from safety boundaries. Because the entire pipeline—including the barrier conditions—is differentiable, standard machine learning tools like stochastic gradient descent can optimize the controller for both performance and safety simultaneously.
Experimental Results
The method was evaluated on a simulated 1D system tasked with tracking a sinusoidal reference trajectory within a constrained "corridor."
Key results from the simulation include:
- Safety Maintenance: In cases where the DPC controller was improperly trained, the safety filter successfully intervened to "nudge" the system back into the safe corridor.
- Efficiency: For well-trained networks, the safety filter rarely needed to override the policy, allowing the system to achieve MPC-level safety at the near-instantaneous speed of a neural network.
- Robustness: The controller remained safe despite external noise and the discrete-time sampling effects that often destabilize simpler safety mechanisms.
Conclusion and Future Work
This approach demonstrates that neural controllers can be made as safe as traditional optimization-based methods without sacrificing their speed advantage. Current limitations include potential "bang-bang" behavior if buffer zones are too small and a restriction to "relative degree 1" systems. Future research will focus on smoothing control transitions and extending the framework to higher-dimensional, complex nonlinear systems such as multi-rotor drones and robotic manipulators.
Summary updated