Unsupervised Learning Research Papers - Academia.edu (original) (raw)

Center of Mass (CoM) estimation realizes a crucial role in legged locomotion. Most walking pattern generators and real-time gait stabilizers commonly assume that the CoM position and velocity are available for feedback. In this thesis we present one of the first
3D-CoM state estimators for humanoid robot walking. The proposed estimation scheme fuses effectively joint encoder, inertial, and feet pressure measurements with an Extended Kalman Filter (EKF) to accurately estimate the 3D-CoM position, velocity, and external
forces acting on the CoM. Furthermore, it directly considers the presence of uneven terrain and the body’s angular momentum rate and thus effectively couples the frontal with the lateral plane dynamics, without relying on feet Force/Torque (F/T) sensing.
Nevertheless, it is common practice to transform the measurements to a world frame of reference and estimate the CoM with respect to the world frame. Consequently, the robot’s base and support foot pose are mandatory and need to be co-estimated. To this end, we extend a well-established in literature floating mass estimator to account for the
support foot dynamics and fuse kinematic-inertial measurements with the Error State Kalman Filter (ESKF) to appropriately handle the overparametrization of rotations. In such a way, a cascade state estimation scheme consisting of a base and a CoM estimator
is formed and coined State Estimation RObot Walking (SEROW). Additionally, we employ Visual Odometry (VO) and/or LIDAR Odometry (LO) measurements to correct the kinematic drift caused by slippage during walking. Unfortunately, such measurements suffer from outliers in a dynamic environment, since frequently it is assumed that only the
robot is in motion and the world around is static. Thus, we introduce the Robust Gaussian ESKF (RGESKF) to automatically detect and reject outliers without relying on any prior knowledge on measurement distributions or finely tuned thresholds. Therefore, SEROW
is robustified and is suitable for dynamic human environments. In order to reinforce further research endeavors, SEROW is released to the robotic community as an open-source ROS/C++ package.
Up to date control and state estimation schemes readily assume that feet contact status is known a priori. Contact detection is an important and largely unexplored topic in contemporary humanoid robotics research. In this thesis, we elaborate on a broader question: in which gait phase is the robot currently in? To this end, we propose a holistic frame-
work based on unsupervised learning from proprioceptive sensing that accurately and efficiently addresses this problem. More specifically, we robustly detect one of the three gait-phases, namely Left Single Support (LSS), Double Support (DS), and Right Single Support (RSS) utilizing joint encoder, IMU, and F/T measurements. Initially, dimensionality reduction with Principal Components Analysis (PCA) or autoencoders is performed to extract useful features, obtain a compact representation, and reduce the noise. Next, clustering is performed on the low-dimensional latent space with Gaussian Mixture Models (GMMs) and three dense clusters corresponding to the gait-phases are obtained. Interestingly, it is
demonstrated that the gait phase dynamics are low-dimensional which is another indication pointing towards locomotion being a low dimensional skill. Accordingly, given that the proposed framework utilizes measurements from sensors that are commonly available
on humanoids nowadays, we offer the Gait-phase Estimation Module (GEM), an open-source ROS/Python implementation to the robotic community.
SEROW and GEM have been quantitatively and qualitatively assessed in terms of accuracy and efficiency both in simulation and under real-world conditions. Initially, a simulated robot in MATLAB and NASA’s Valkyrie humanoid robot in ROS/Gazebo were employed to establish the proposed schemes with uneven/rough terrain gaits. Subsequently,
the proposed schemes were integrated on a) the small size NAO humanoid robot v4.0 and b) the adult size WALK-MAN v2.0 for experimental validation. With NAO, SEROW was implemented on the robot to provide the necessary feedback for motion planning and real-
time gait stabilization to achieve omni-directional locomotion even on outdoor/uneven terrains. Additionally, SEROW was used in footstep planning and also in Visual SLAM with the same robot. Regarding WALK-MAN v2.0, SEROW was executed onboard with kinematic-inertial and F/T data to provide base and CoM feedback in real-time. Furthermore, VO has also been considered to correct the kinematic drift while walking and facilitate possible footstep planning. GEM was also employed to estimate the gait phase in WALK-MAN’s dynamic gaits.
Summarizing, a robust nonlinear state estimator is proposed for humanoid robot walking. Nevertheless, this scheme can be readily extended to other type of legged robots such
as quadrupeds, since they share the same fundamental principles.