Real-Time Embedded Robotic Control Using Stability-Constrained Neuro-Fuzzy Reinforcement Learning

Real-time robotic control under uncertainty remains challenging due to nonlinear dynamics, external disturbances, and the embedded hardware's stringent computational and memory constraints. While fuzzy logic and adaptive neuro-fuzzy controllers provide robustness and interpretability, their performance degrades without systematic adaptation, and reinforcement learning-based controllers often remain impractical for microcontroller deployment due to stability and real-time feasibility concerns. This paper presents a hybrid adaptive neuro-fuzzy inference system enhanced with reinforcement learning and Lyapunov-based stability supervision for real-time embedded robotic control. The proposed controller employs a compact first-order Sugeno ANFIS policy trained online using an actor-critic reinforcement learning framework, where a Lyapunov condition constrains parameter updates to enforce stable closed-loop operation during learning. The controller is implemented using fixed-point arithmetic, quantization-aware learning, and rule-base pruning within a TinyML-oriented embedded architecture to enable deployment on low-cost microcontrollers. The proposed approach is implemented on STM32F407 and ESP32-S3 microcontrollers and experimentally validated on two physical robotic platforms: a differential-drive mobile robot and a 3degree-of-freedom robotic manipulator. Across repeated experimental trials, the proposed controller reduces mobile robot heading overshoot from approximately 10-12% with PID and 5-6% with static fuzzy control to approximately 1%, while shortening settling time by up to 35% relative to fuzzy control and reducing RMS path-tracking error under dynamic obstacle scenarios by approximately 15-20%. Integrated navigation energy consumption is reduced by up to 7% compared with that of PID control. On the manipulator platform, the controller achieves an overshoot below 1° across all joints under both unloaded and payload conditions, compared with an overshoot exceeding 10° for PID under payload, while reducing the RMS tracking error by more than 80% relative to PID and by 20-30% relative to static fuzzy control. Disturbance rejection experiments show reductions exceeding 50% in peak deviation and recovery time compared with PID. Embedded execution measurements demonstrate real-time feasibility with sub millisecond control-loop latency, CPU use below 25%, and total memory usage under 60 kB. The results demonstrate that stability-constrained neuro-fuzzy reinforcement learning can be executed directly on embedded robotic hardware, providing a practical balance between adaptability, robustness, and real-time performance.