Learning Robust Task Priorities and Gains for Control of Redundant Robots

Abstract

Generating complex movements in redundant robots like humanoids is usually done by means of multitask controllers based on quadratic programming, where a multitude of tasks is organized according to strict or soft priorities. Time-consuming tuning and expertise are required to choose suitable task priorities, and to optimize their gains. Here, we automatically learn the controller configuration (soft and strict task priorities and Convergence Gains), looking for solutions that track a variety of desired task trajectories efficiently while preserving the robot's balance. We use multiobjective optimization to compare and choose among Paretooptimal solutions that represent a trade-off of performance and robustness and can be transferred onto the real robot. We experimentally validate our method by learning a control configuration for the iCub humanoid, to perform different whole-body tasks, such as picking up objects, reaching and opening doors.

Prescient teleoperation of humanoid robots

“…Each of these trajectories has a different priority which determines how the robot's controller executes the entire movement (Methods): the top priority is given to the center of mass (to avoid falling) and the feet poses (which should not move in our experiments, since we only target double support motions), and the postural trajectories have the lowest priority. The exact hard and soft priorities were found in previous work with a multi-objective stochastic optimizer so that the robot is unlikely to fall but tracks the trajectories as well as possible (5).…”

“…The whole-body motion of the operator is captured with a motion capture suit (Methods), but it cannot be directly used as reference for the robot because of the difference in kinematics (e.g., joint limits, body dimensions) and dynamics (e.g., mass distribution). The system therefore needs to "retarget" the motion (4,5,6), that is, to compute references that make sense for the robot. To do so, Cartesian references are scaled using a fixed factor that accounts for the size difference between the human and the operator (Methods).…”

“…Each sample of each of these trajectories represents a control reference ŷ. Given ŷ, the robot commands q are generated by solving the redundant inverse kinematics, which can be formulated as a constrained quadratic programming problem with equality and inequality constraints (17,5):…”

“…Two main challenges are raised by humanoid teleoperation: (i) how to map a whole-body motion of the human to a robot that has close, but different physical constraints and dynamics (4,5), and (ii) how to cope with the unavoidable communication delays between the movement of the operator and the feedback. The present paper addresses both of these challenges but focuses on delay compensation because substantial work has already been achieved for whole-body retargeting (4,6).…”

Evolving the Behavior of Machines: From Micro to Macroevolution

“…Take the limits of the simulation into account by learning to predict which controllers will transfer well (Koos et al, 2012); Evolve many different solutions with a quality diversity (Cully et al, 2015) or multi-objective algorithm (Penco et al, 2020), then test on the real system to find those that work best on the real robot.…”