David Remy at ETH Zurich has performing his graduation research on the stabilization of passive dynamic walking quadrupeds.
Related Publications:
Remy, C.D., Buffinton, K.W., and Siegwart, R.Y., 2010, "Stability Analysis of Passive Dynamic Walking of Quadrupeds," International Journal of Robotics Research. (in press).
Abstract
We introduce a detailed numerical simulation and analysis framework
to extend the principles of passive dynamic walking to quadrupedal
locomotion. Non-linear limit cycle methods are used to identify possible
gaits and to analyze the stability and efficiency of quadrupedal
passive dynamic walking. In doing so, special attention is paid to
issues that are inherent to quadrupedal locomotion, such as the occurrence
of simultaneous contact collisions and the implications of
the phase difference between front and back leg pairs. Limit cycles
identified within this framework correspond to periodic gaits and
can be placed into two categories: in-phase gaits in which front and
back legs hit the ground at roughly the same time, and out-of-phase
gaits with a 90 phase shift between the back and front leg pairs.
The latter are, in comparison, energetically more efficient but exhibit
one unstable eigenvalue that leads to a phase divergence and
results in a gait-transition to a less efficient in-phase gait. A detailed
analysis examines the influence of various parameters on stability
and locomotion speed, with the ultimate goal of determining a stable
solution for the energy-efficient, out-of-phase gait. This was achieved
through the use of a wobbling mass, i.e. an additional mass that is
elastically attached to the main body of the quadruped. The methods,
results, and gaits presented in this paper additionally provide a point
of departure for the exploration of the considerably richer range of
quadrupedal locomotion found in nature.
Research Abstract - Master's Student Christian Hubicki
The field of legged robotics has been long anticipated in the popular media to herald a revolution in both civilian and military life. From mechanical fire fighters barreling through burning apartments with minimal regard for self-preservation to nimble explorers bounding up Martian ridges who never complain about the cold, finding applications for bipedal machines requires little imagination. Despite their promised dexterity and overall popular appeal, in the early 21st century, bipedal robots are seldom sighted outside of university research labs or cutting-edge technology firms.
The absence of these legged machines in our daily lives can be attributed to significant technical barriers in performance. The largely untold flaw of Honda’s flagship robotic humanoid, ASIMO, is that its exorbitant energy consumption drains its generously sized battery pack in roughly 30 minutes, nullifying its utility outside of relatively short public demonstrations. Recognizing that this energy limitation is not unique to ASIMO but common among current-generation walking robots, academic researchers have recently pushed to develop highly efficient bipeds. The consequence was a series of prototypes which trade an abundance of actuation and control authority for an under-actuated approach dubbed Dynamic Walking. Specifically, Cornell University developed two internationally publicized walking machines; one which boasted energy efficiency on par with human walking (for short distances) and the Cornell Ranger which set a world record for walking 5.6 miles on a single battery charge.
While delivering such significant advances in energy efficiency, dynamic walking robots have still largely fallen short in applications with high speed requirements or rough terrain. This investigation uses simulation to explore the inherent tradeoffs of controlling high-speed and highly robust walking robots while minimizing energy consumption. Using a novel controller which optimizes robustness, energy efficiency, and speed of a simulated robot on rough terrain, the user can adjust their priorities between these three outcome measures and systematically generate a characteristic performance curve. This curve represents the entire spectrum of performance for the given robot, revealing necessary energy costs for various demands of speed and robustness.
The novel robot controller is a two-tiered hierarchical system consisting of a heuristically-driven single-step controller and an overseeing Artificial Intelligence algorithm. The single-step controller is generated by tuning a set of control parameters in order to closely approximate an optimal performance curve produced by a series of genetic optimizations. This single step is calculated to produce the desired step speed for the least amount of energy expenditure. The Artificial Intelligence algorithm, more precisely described as a Value Iteration Reinforcement Learning Algorithm, is tasked to optimally plan for future steps. This step-planning algorithm decides which actions are taken by the single-step controller, thereby seeking conditions conducive to superior walking performance and avoiding unfavorable situations which the single-step controller lacks the foresight to evade.
2009 Hubicki Research Symposium Poster
2010 Hubicki Research Symposium Poster