Adaptable Recovery Behaviors in Robotics: A Behavior Trees and Motion Generators(BTMG) Approach for Failure Management

TL;DR

The paper tackles robust failure management in dynamic collaborative robots by extending the Behavior Trees and Motion Generators (BTMG) framework with adaptable recovery behaviors whose extrinsic parameters are learned through reinforcement learning (RL). It formalizes recovery behaviors as skilled primitives and leverages Bayesian optimization to tune parameters, enabling on-the-fly adaptation during task execution, as evidenced by peg-in-a-hole experiments with a dual-arm KUKA iiwa. A multi-objective evaluation highlights improvements in task success and efficiency while balancing forces exerted by the end-effector, validating the approach's resilience under varied failures. Overall, the work contributes a RL-guided extension to BTMG for adaptive recovery in robotics and outlines future directions for automated recovery pipelines and diagnostic capabilities using BT tick signals, with significant implications for real-time adaptability in collaborative systems.

Abstract

In dynamic operational environments, particularly in collaborative robotics, the inevitability of failures necessitates robust and adaptable recovery strategies. Traditional automated recovery strategies, while effective for predefined scenarios, often lack the flexibility required for on-the-fly task management and adaptation to expected failures. Addressing this gap, we propose a novel approach that models recovery behaviors as adaptable robotic skills, leveraging the Behavior Trees and Motion Generators~(BTMG) framework for policy representation. This approach distinguishes itself by employing reinforcement learning~(RL) to dynamically refine recovery behavior parameters, enabling a tailored response to a wide array of failure scenarios with minimal human intervention. We assess our methodology through a series of progressively challenging scenarios within a peg-in-a-hole task, demonstrating the approach's effectiveness in enhancing operational efficiency and task success rates in collaborative robotics settings. We validate our approach using a dual-arm KUKA robot.

Figures (8)

• Figure 1: The real dual-arm KUKA iiwa setup executing a handover task before inserting the peg in the orange block. On the table, the purple peg for Scenario 1, alongside various obstacles that block the opening an be seen.
• Figure 2: The figure shows the peg-in-a-hole task execution using our approach. We have separate sets of production and recovery behaviors. We can use a planner to come up with a sequence for a given failure specification. Subsequently, we tune the learnable parameters via reinforcement learning. Ultimately, appropriate recovery behaviors and skills are applied based on the identified failure, ensuring successful peg insertion. In the image we see successful peg insertions for all the scenarios.
• Figure 3: Illustration of the PegInsertion skill alongside its associated failure scenarios and corresponding recovery behaviors. Panel (a) illustrates the initial and final states of Scenario 1. Panel (b) displays the initial state of Scenario 2 with the recovery behavior of pick-place. Panel (c) showcases the initial state of Scenario 3 with the recovery behavior of push. Lastly, panel (d) presents the initial states of Scenarios 4 and 5 with the recovery behavior of pick-exchange.
• Figure 4: Pareto front for Scenario 1: Baseline. Each experiment is denoted by a distinct color, with each bold point representing a pareto-optimal policy ready for execution. The optimizer tries to strike a balance between the reward for successful insertion and the force applied by the end-effector.
• Figure 5: Pareto front for Scenario 2: Static Recovery. This demonstrates that achieving a higher insertion reward necessitates greater force application, as observed from the force exerted by the end-effector during the search for the hole.