| Achoimre: | <p>Quadrupedal robots are designed to offer efficient and robust mobility on uneven terrain. This thesis investigates combining numerical optimization and machine learning methods to achieve interpretable kinodynamic planning of natural and agile locomotion.</p>
<p>The proposed algorithm, called Receding-Horizon Experience-Controlled Adaptive Legged Locomotion (RHECALL), uses nonlinear programming (NLP) with learned initialization to produce long-horizon, high-fidelity, terrain-aware, whole-body trajectories. RHECALL has been implemented and validated on the ANYbotics ANYmal B and C quadrupeds on complex terrain.</p>
<p>The proposed optimal control problem formulation uses the single-rigid-body dynamics (SRBD) model and adopts a direct collocation transcription method which enables the discovery of aperiodic contact sequences. To generate reliable trajectories, we propose fast-to-compute analytical costs that leverage the discretization and terrain-dependent kinematic constraints.</p>
<p>To extend the formulation to receding-horizon planning, we propose a segmentation approach with asynchronous centre of mass (COM) and end-effector timings and a heuristic initialization scheme which reuses the previous solution. We integrate real-time 2.5D perception data for online foothold selection. Additionally, we demonstrate that a learned stability criterion can be incorporated into the planning framework.</p>
<p>To accelerate the convergence of the NLP solver to locally optimal solutions, we propose data-driven initialization schemes trained using supervised and unsupervised behaviour cloning. We demonstrate the computational advantage of the schemes and the ability to leverage latent space to reconstruct dynamic segments of plans which are several seconds long.</p>
<p>Finally, in order to apply RHECALL to quadrupeds with significant leg inertias, we derive the more accurate lump leg single-rigid-body dynamics (LL-SRBD) and centroidal dynamics (CD) models and their first-order partial derivatives. To facilitate intuitive usage of costs, constraints and initializations, we parameterize these models by Euclidean-space variables. We show the models have the ability to shape rotational inertia of the robot which offers potential to further improve agility.</p>
|
|---|