Learning Stack-of-Tasks Management for Redundant Robots

TL;DR

This work tackles redundancy management in high-DOF robots by automatically learning complete Stack-of-Tasks structures directly from a user-defined cost. It introduces Genetic Programming guided by sim-to-real episodic evaluation to optimize the discrete task ordering, activation flags, and continuous gains, yielding interpretable, expert-like hierarchies without manual tuning. The approach demonstrates robust sim-to-real transfer on a dual-arm mobile robot (mobile-YuMi), with reliable obstacle avoidance and high tracking accuracy in static and dynamic environments, even in the presence of distractors. Overall, the framework offers a scalable, user-driven alternative to hand-crafted SoT design, capable of adapting task hierarchies to multi-objective trade-offs while maintaining safety and performance.

Abstract

This paper presents a novel framework for automatically learning complete Stack-of-Tasks (SoT) controllers for redundant robotic systems, including task priorities, activation logic, and control parameters. Unlike classical SoT pipelines-where task hierarchies are manually defined and tuned-our approach optimizes the full SoT structure directly from a user-specified cost function encoding intuitive preferences such as safety, precision, manipulability, or execution speed. The method combines Genetic Programming with simulation-based evaluation to explore both discrete (priority order, task activation) and continuous (gains, trajectory durations) components of the controller. We validate the framework on a dual-arm mobile manipulator (the ABB mobile-YuMi research platform), demonstrating robust convergence across multiple cost definitions, automatic suppression of irrelevant tasks, and strong resilience to distractors. Learned SoTs exhibit expert-like hierarchical structure and adapt naturally to multi-objective trade-offs. Crucially, all controllers transfer from Gazebo simulation to the real robot, achieving safe and precise motion without additional tuning. Experiments in static and dynamic environments show reliable obstacle avoidance, high tracking accuracy, and predictable behavior in the presence of humans. The proposed method provides an interpretable and scalable alternative to manual SoT design, enabling rapid, user-driven generation of task execution hierarchies for complex robotic systems.

Figures (10)

• Figure 1: Overview of the proposed framework. Candidate Stack-of-Tasks (SoT) controllers are first evaluated in simulation to compute a task-specific cost. A Genetic Programming process selects and recombines promising SoTs to iteratively generate improved candidates. The best evolved SoT is subsequently deployed on the ABB YuMi mobile manipulator for experimental validation.
• Figure 2: Schematic of a three-task Stack of Tasks (SoT) null space projection. Each task contributes to the final joint velocity command through its own projected reference, forming the hierarchical control law $\dot{\mathbf{q}}=\dot{\mathbf{q}}_{1}+N_1\dot{\mathbf{q}}_{2}+N_{12}\dot{\mathbf{q}}_{3}$.
• Figure 3: Example of the robot performing an Inverse Kinematic sequence.
• Figure 4: Graphical User Interface (left) for configuring learning simulation (right) parameters. The interface allows users to define the relative weights of cost criteria (Accuracy, Safety, Manipulability, Speed) and set simulation parameters, including population size, number of iterations, and simulation duration. Designed for ease of use, it supports intuitive navigation with dedicated controls for starting or exiting the simulation.
• Figure 5: Example with the result of the % of individuals in a generation that have the same priority order as the final solution, in different learning runs. The base only case is reported on the left, in which the order is [OA, IK]. On the right, the result obtained with the right arm + base, in which the order is [OA, IK, Max. Manip., Max MJL].