Dynamic quadruped locomotion implies high-intensity integration toward environmental factors and requires considering the information from sensory feedback. The authors represent CPG-based locomotion model with sensorimotor coordination. They build an efficient integration between motor and sensory neurons that can generate dynamic behavior, especially in locomotion coordination during leg malfunction. They emphasize an optimization strategy to optimize the interconnection structure of CPG-based locomotion model. They use a multi-objective evolutionary algorithm to optimize the synaptic weight between motor–motor neurons and motor–sensory neurons. The applied cascade optimization is 1) dynamic gait pattern optimization using desired speed and torso oscillation as the fitness function and 2) malfunction compensation optimization using moving direction error and torso oscillation as the fitness evaluation. The proposed model has been applied to simulated and real middle-size quadruped robots. It showed the proposed optimization can generate a smooth transition during a robot's leg unction.
TopIntroduction
The development of legged robots has increased significantly. Several researchers focus on the development of locomotion systems. They try to find how to achieve dynamic locomotion and how to transfer the locomotion model of humans or animals to legged robots. Naturally, the locomotion model of humans or animals integrates several aspects, such as sensing, motion generation, perception, and stability systems (Patla, 1997). Most researchers apply some constraints to simplify the role of several aspects in the locomotion system. It simplifies the integration of the cognition system and behavior systems. However, this simplifying strategy falls short of achieving dynamic locomotion. The simplified model has been applied from biological, mechanical, and cognitive perspectives (Clark and Grush, 1999; Hinton et al., 2012; Saputra, Takeda, et al., 2015; Saputra, Tay et al., 2016).
On the other hand, human and animal movements are generated through the role of sensorimotor coordination. It builds an integration system between sensing and acting through motor neurons (Frost et al., 2015). Sensorimotor coordination integrates interoceptive and exteroceptive sensory systems to generate a motion system and orientation system. Then, the visual system is integrated with other sensory systems to play a role in the musculoskeletal systems (Forbush et al., 2008). For example, when a person is walking and encounters a sudden obstacle during the swinging phase, the exteroceptive sensory information triggers the joint actuator to lift and avoid the obstacle. The visual input from the retina is transferred to the cerebellum, basal ganglia, and supplementary motor area. The central nervous system will combine several sensory inputs to avoid inaccurate information from a single sensory input (Kisner and Colby, 2002). When some legs are broken, certain animals generate new behaviors to make their movements more efficient. In dynamic locomotion, the generated movement has to have an objective.
Dynamic locomotion implies a compromise of interdisciplinary studies, especially from a neurobiological model. To realize a neurobiological model, cognition, embodiment structure, and locomotion generator should be integrated (Pfeifer and Bongard, 2006). Thus, it becomes possible to build an advanced locomotion system that deals with dynamic, internal, and external sensory information. The changing of embodiment structure, such as leg malfunction, can be addressed by integrating embodiment structure and locomotion generator.
To solve this issue, some reserchers try realizing dynamic locomotion using a neural-based approach (Ijspeert and Cabelguen, 2006). It widely used in quadruped robot locomotion (Liu et al., 2013; Saputra, Tay, et al., 2016). Previous studies focused on neural models for motion generation on rough terrain Kimura et al., 2007 and speed-dependent motion patterns (Maufroy et al., 2008). Neural systems enable transitioning between movement patterns and adjusting gait for energy efficiency (Aoi et al., 2013; Owaki and Ishiguro, 2017). Our prior research proposed a neural-based locomotion planner utilizing body posture and contact points as feedback signals (Saputra, Botzheim, et al., 2016; Saputra and Kubota, 2017; Saputra et al., 2018). However it, is difficult to realize integration of sensory feedback with leg malfunction systems. Furthermore, single objective evolutionary algorithm has difficulties to realize multiple objective in dynamic locomotion. We posed three scientific issues based on the state of the art:
- •
How to build the locomotion model with considering sensory feedback and leg malfunction
- •
How to apply multi-objective optimization in the above locomotion model?
- •
How to optimize the locomotion model for quadruped robot?
To solve the above questions, we applied a neural-based model proposed in the previous model (Saputra, Ijspeert, et al., 2020) and use multi-objective evolutionary algorithms to optimize multiple objective functions. Then, we use computer simulation to optimize the model before transferring it to the real robot. The detail background and related work of neural based locomotion and optimization model used in locomotion can be seen in the following subsection.
This chapter is organized as follows: the proposed CPG model is presented in Section II; the multi objective evolutionary algorithm is delineated in Section III; Section IV presents the experimental results; finally, in Section V, we conclude the research and discuss the prospects of the proposed model.