Learning Active Spine Behaviors for Dynamic and Efficient Locomotion in Quadruped Robots

Abstract

In this work, we provide a simulation framework to perform systematic studies on the effects of spinal joint compliance and actuation on bounding performance of a 16-DOF quadruped spined robot Stoch 2. Fast quadrupedal locomotion with active spine is an extremely hard problem, and involves a complex coordination between the various degrees of freedom. Therefore, past attempts at addressing this problem have not seen much success. Deep-Reinforcement Learning seems to be a promising approach, after its recent success in a variety of robot platforms, and the goal of this paper is to use this approach to realize the aforementioned behaviors. With this learning framework, the robot reached a bounding speed of 2.1 m/s with a maximum Froude number of 2. Simulation results also show that use of active spine, indeed, increased the stride length, improved the cost of transport, and also reduced the natural frequency to more realistic values.

On the optimal spine morphology of rapidly accelerating quadrupeds

“…This was achieved using Contact-Implicit Optimization methods [36], which have been shown to be a vital tool in studying locomotion [37]- [39], and has been implemented with significantly increased accuracy [34]. This method requires a large number of complementary constraints which can be found in [36] equation (8) to (16). All contacts are modelled as an inelastic collision [36], [40] and can take one of two modes (sticking or slipping).…”

“…These spines often consist of a multitude of actuated rigid links. Despite being common place and crucial in nature, the spine (active [9]- [16] or passive [17]- [21]) is not commonly found in robotics. To date there Fig.…”

“…Statistical analysis was performed on these results to determine the optimal spine morphology. has been little focus on transient maneuvers in robotics and the main focus has been on steady-state locomotion and energy efficiency [11], [22] often utilizing the spines as a passive actuator (a spring system [18], [20]) as opposed to an active spine [10], [12]- [16]. This may be one of the reasons as to why the current robotic platforms are not as agile as animals.…”

Gait and Trajectory Optimization by Self-Learning for Quadrupedal Robots with an Active Back Joint

“…We are interested in a smooth trajectory that requires minimal power consumption while maximizing the resulting velocity. e efficiency is the ratio between the input power and the power consumption (i.e., the ratio between the average kinetic energy ∼ v 2 and the input power W, compare with [29]). We also include an additional dimensionless ratio d/T which measures the "straightness" of the path (compare with [37]).…”

“…Obviously, using cameras as sensors in the gait optimizing procedure requires real-time techniques for analyzing the video stream (see [27,28]). e back joint (spine) was investigated in [29], where the research studies presented simulations showing that an active back is a key driving factor for the improvement of speed and cost-of-transport in quadrupeds. In [30], Khoramshahi et al presented a simplified (wheeled) locomotion system which is behaviorally and structurally similar to a galloping quadruped and showed that fast locomotion requires a flexible spine.…”