Ευφυής έλεγχος ρομποτικού χεριού για επιδέξιο χειρισμό αντικειμένων

This thesis addresses in-hand cube reorientation, a benchmark task in dexterous robotic manipulation, on the low-cost anthropomorphic LEAP HAND v1 Lite. The objective is to reorient a cube held in the palm toward a sequence of relative single-axis rotation targets, fixed at 90° or randomized between 45° and 135°, by continuously replanning the finger motions, both in a MuJoCo physics simulation and on the physical hand. The core of the system is a sampling-based Model Predictive Path Integral (MPPI) controller. On top of a baseline MPPI formulation, four extensions from the sampling-based control literature are adopted and combined to improve sampling efficiency and control smoothness: antithetic sampling for variance reduction, an adaptive temperature parameter, a spline-parameterized control sequence and adaptive noise standard deviation σ. The controller optimizes a cost function that combines an orientation-error term, a keep-in-palm-term, and control effort and smoothness regularization, and runs in real time at 25 Hz. To transfer the controller to hardware, a vision-based perception pipeline estimates the 6-DoF pose of the cube from multiple AprilTags using PnP solving and an error-state quaternion Kalman filter, while observation noise and physics randomization are used to assess the controller’s robustness to model mismatch. The method is evaluated in simulation against a vanilla MPPI baseline and on the real hand, using an evaluation protocol based on the interquartile mean and bootstrap confidence intervals.