Sustaining the gait locomotion in an adversarial environment requires the robot to react to novel experiences adaptively. In Free Energy Principle (FEP), the behavioral reaction is driven by the discrepancy between observation and prediction. Although, for legged robot gait locomotion, the prediction of gait dynamics is challenging as the consequences non-linearly depend on the activity history, the animal gait is robust, adapting to severe motion disruptions seemingly instantly. In biomimetic robotics, the Central Pattern Generator (CPG) relaxes the general dynamics of body-environment interaction to the stable and repetitive dynamics of gait. Based on these observations, we propose self-learning of the gait dynamics model and FEP framework that infers state estimation and gait control. The proposed method is experimentally evaluated on a real hexapod walking robot with 18 controllable degrees of freedom. The robot learns the gait dynamics model indoors and then deploys it in outdoor navigation under various adversarial scenarios. Results show that the developed interpretable gait controller exhibits complex and real-time adaptive behavior when it encounters unknown situations.

Sustaining the robot's longevity becomes challenging in dynamic deployments characterized by new unknown environments and embodiments outside of the prior knowledge. Hence, the knowledge of robot-environment interactions needs to be continually updated for system adaptation. It can be implemented through self-verification as a continual comparison of predictions with observations using the predictive coding (PC) principle. The principle has been further extended into the active inference control (AIC) in biomimetic robotics to drive the control, state estimation, and model update. However, continually updating one model leads to catastrophic forgetting in the long term. Therefore, we propose an autonomously expanding self-verifying world model (WM) of sensorimotor dynamics utilized in model-based gait control. The model combines PC with the incremental knowledge representation based on the internal model (IM) principle. The proposed method is experimentally validated in virtual and real scenarios, where the hexapod walking robot has to recognize and adapt to leg paralysis and then recognize the recovery. The method generates novel behaviors in real time, improving the performance and outperforming the examined state-of-the-art methods. Furthermore, the robot's decisions and gained knowledge are interpretable and promise further functional scalability.

In this paper, we propose a novel dynamic gait controller for the repetitive behavior of soft robot manipulators performing routine tasks. Compliance with soft robots is advantageous when the robot interacts with living organisms and other fragile objects. However, predicting and controlling repetitive behavior is challenging because of hysteresis and non-linear dynamics governing the interactions. Existing priorfree methods track the dynamic state using recurrent neural networks or rely on known generalized coordinates describing the robot's state. We propose to model the interaction induced by the repetitive behavior as gait dynamics and represent the dynamic state with Central Pattern Generator (CPG) tracking the motion phase and thus reduce the complexity of the robot's forward model. The proposed method bootstraps an ensemble of the forward models exploring multiple dynamic contexts that are expanded as it searches for repetitive motion producing the target repetitive behavior. The proposed approach is experimentally validated on a pneumatically actuated soft robot arm I-Support, where the method infers gaits for different targets.

In this work-in-progress report, we present experimental results of lightweight learning-based leg-contact detection methods for a small hexapod walking robot with position feed- back only. The detection of the leg contact with the surface is addressed as anomaly detection using predicted and measured positions of the leg’s joints in the leg swing phase. A polynomial regressor and three-layer neural network are evaluated regarding the prediction error and computational requirements using realistic datasets collected with the real hexapod walking robot.

Motivated by observations of animal behavior, locomotion of multi-legged walking robots can be controlled by the central pattern generators (CPGs) that produce a repetitive motion pattern. A rhythmic pattern, a gait, is defined by phase relations between all leg joints. In a case of an external influence such as terrain irregularity, some actuator phase can shift and thus disrupt the phase relations between the actuators. The actuator phase relations can be maintained only by synchronizing to the sensors, which output can indicate the motion disruption. However, establishing correct sensory-motor phase relations requires not only the motor phase model but also a model of the sensory phase, which is generally unknown. Although both sensory and motor phases can be modeled by single CPG, the capabilities of such CPG-based controllers are limited because they are not flexible and robust. In this paper, we propose to model the phases of each sensor and motor by separate CPGs. The phase relations between the sensor and motor phases are established by radial basis function (RBF) neurons learned with proposed periodic Grossberg rule for which we present the convergence proof. Based on the reported evaluation results using high-fidelity simulation, the proposed locomotion controller demonstrates the desired plasticity, and it is capable of learning multiple gaits with robust synchronization to terrain changes using sensor inputs.