Abstract
In this thesis, a biped which combines the merits of both powered and passive bipeds is proposed to walk in semi-active manner. The semi-activeness is motivated by Tai-Chi, the ancient Chinese martial art, in which the practitioner has to learn to actuate only part of his/her muscles to control the body posture and relax the others. Likewise, for the proposed biped, during most of a walking cycle, only half of the joints are actuated to follow specific trajectories, and the other half remain unactuated but have passive springs connected between adjacent links. It is expected that by having unactuated joints, the biped can preserve the power saving feature of the passive biped, and by having actuated joints under active control, the biped can also achieve the stability and performance of the powered biped. To devise a systematic design methodology for the biped, its dynamics as well as the walking constraints are carefully studied. Furthermore, an optimization procedure is also proposed to compute the optimal trajectories for the actuated joints and spring constants which can lead to minimum energy consumption. The feasibility of the proposed biped, including the system design and the control strategy, is verified by numerical simulations and hardware implementation. Experiments indicate that the biped walking in the semi-active manner consumes 80% less the electrical power of the fully-powered biped that performs the same gait and is more energy-efficient than several state-of-the-art bipeds. Due to experimental effects such as unmodelled dynamics, measurement noise, and so on, the optimized actuation trajectories are still relied on trial-and-error tuning to make the biped achieve satisfactory walking performance. An iterative learning strategy is thus proposed to provide further experimental tuning on the actuated joint trajectories computed by the model-based optimization procedure. The strategy demands the biped to swing its leg repeatedly, and during each swinging the hip trajectory is iteratively modified by a learning law to minimize foot scuffing of the swing leg and yet keep the associated foot clearance to within a small limit for reducing power consumption. Experiments show that the learning strategy leads to a convergent hip trajectory after 11 iterations. Besides, when the trajectory is adopted for actual walking, the biped demonstrates a good balance between power efficiency and robustness to ground conditions.