| Abstract |
Legged robot locomotion in unstructured and slippery terrains relies heavily on
accurately identifying the contact state between the robot’s feet and the ground.
Contact state estimation poses significant challenges, typically addressed by leveraging force measurements, joint encoders as well as robot kinematics and dynamics. This thesis introduces two novel approaches for accurately estimating the
contact state in real-time, namely, a deep learning approach and a probabilistic
model-based method.
To address the challenges of leg contact detection in bipedal walking gaits, a
deep learning framework is proposed. This framework accurately and robustly
estimates the contact state probability for each leg, distinguishing between stable
contact, slip, or no contact. Notably, the framework relies solely on proprioceptive sensing and demonstrates generalizability across diverse friction surfaces and
legged robotic platforms. Comprehensive evaluations, including comparisons with
state-of-the-art methods, have been performed using ATLAS, NAO, and TALOS
humanoid robots. Furthermore, the framework’s efficacy is demonstrated in realworld base estimation tasks with a TALOS humanoid robot.
The second proposed approach is model-based and relies solely on Inertial
Measurement Units (IMUs) mounted on the robot’s end effectors. It offers a versatile approach that can be implemented in any legged robot without the necessity
of training data. By capitalizing on the uncertainty of IMU measurements, this
novel probabilistic method is capable of estimating the probability of stable contact. The method approximates the multimodal probability density function using
Kernel Density Estimation, providing reliable contact state estimation. Extensive
evaluations of the proposed method have been conducted on both real and simulated scenarios, demonstrating its effectiveness on various bipedal and quadrupedal
robotic platforms, including ATLAS, TALOS, and Unitree’s GO1.
Finally, this thesis introduces an application of the aforementioned probabilistic
contact state estimator that further demonstrates its efficacy. More specifically, an
adaptive trajectory tracking controller is presented, which was developed by peers
in the Computational Vision and Robotics Laboratory. This controller consists of
two prioritized layers of adaptation aimed at preventing leg slippage when stepping
on partially or globally slippery terrains. The primary emphasis is placed on the
results, as the first layer of adaptation effectively utilizes the contact probability
to distribute the effort among each leg. Therefore, the accuracy of this controller
is directly correlated to the ability to estimate the contact state in real-time which
validates the robustness of the proposed contact estimator.
|
Collections
