A dynamically-Balanced Walking Biped

Describes the mechanical, electronic and software design of a 10-DOF bipedal robot which has been constructed to study control, parameterisation and automatic expansion of the stability envelope of a complex real-time behaviour, namely, dynamically-balanced two-legged walking. The machine is physically complete and demonstrates reasonable reliability in movement control including dynamically-balanced standing. High-level reinforcement learning code is being developed to extend this to walking. The machine offers a challenging problem domain to the flourishing machine learning community and represents a shift in emphasis, away from learning algorithms that work on simplified, preprocessed, artificial and static data sets to learning heuristics which deal with noisy, real-time data collected from sensors on a dynamic, real-world system. A Dynamically-Balanced Walking Biped Graham Mann † Bruce Armstrong † Phil Preston ‡ Barry Drake ‡ Abstract. Describes the mechanical, electronic and software design of a 10DOF bipedal robot which has been constructed to study control, parameterisation and automatic expansion of the stability envelope of a complex real-time behaviour, namely, dynamically-balanced two-legged walking. The machine is physically complete and demonstrates reasonable reliability in movement control including dynamically-balanced standing. High-level reinforcement learning code is being developed to extend this to walking. The machine offers a challenging problem domain to the flourishing machine learning community and represents a shift in emphasis, away from learning algorithms that work on simplified, preprocessed, artificial and static data sets to learning heuristics which deal with noisy, real-time data collected from sensors on a dynamic, real-world system. Describes the mechanical, electronic and software design of a 10DOF bipedal robot which has been constructed to study control, parameterisation and automatic expansion of the stability envelope of a complex real-time behaviour, namely, dynamically-balanced two-legged walking. The machine is physically complete and demonstrates reasonable reliability in movement control including dynamically-balanced standing. High-level reinforcement learning code is being developed to extend this to walking. The machine offers a challenging problem domain to the flourishing machine learning community and represents a shift in emphasis, away from learning algorithms that work on simplified, preprocessed, artificial and static data sets to learning heuristics which deal with noisy, real-time data collected from sensors on a dynamic, real-world system.