A Method for Designing Autonomous Robots that Know Their Limits

Abstract

For assisting humans in their daily lives, robots need to perform long-horizon tasks, such as tidying up a room or preparing a meal. One effective strategy for handling a long-horizon task is to break it down into short-horizon subgoals, that the robot can execute sequentially. In this paper, we propose extending a predictive learning model using deep neural networks (DNN) with a Subgoal Proposal Module (SPM), with the goal of making such tasks realizable. We evaluate our proposed model in a case-study of a long-horizon task, consisting of cutting and arranging a pizza. This task requires the robot to consider: (1) the order of the subtasks, (2) multiple subtask selection, (3) coordination of dual-arm, and (4) variations within a subtask. The results confirm that the model is able to generalize motion generation to unseen tools and objects arrangement combinations. Furthermore, it significantly reduces the prediction error of the generated motions compared to without the proposed SPM. Finally, we validate the generated motions on the dual-arm robot Nextage Open.

AlegAATr the Bandit

“…In this paper, we introduced a new bandit algorithm called AlegAATr. AlegAATr leverages Assumption-Alignment Tracking (AAT), a technique proposed in the robotics literature to perform proficiency self-assessment [14,6], to predict the performance of each of the N behaviors that are available to it. It then uses these predictions to select generators at any given time.…”

“…Before defining AlegAATr, we review Assumption-Alignment Tracking (AAT) [14,6], which AlegAATr uses to predict the performance of a generator G at any given time. At each time step t, the generator G takes as input an encoding of the state of the world, denoted s G t , and outputs some action a t for the agent to execute in that time step.…”

“…As a step toward overcoming these limitations, we introduce a new bandit algorithm, called AlegAATr (pronounced al@gad@r). Ale-gAATr leverages Assumption-Alignment Tracking (AAT) [14,6], a technique recently introduced in the robotics literature, to predict the future performance of each of the agent's behavior generators at any given time. In AAT, the agent tracks the veracity of the various assumptions that a generator relies on, and in this way forms a veracity vector.…”

Event-triggered robot self-assessment to aid in autonomy adjustment

“…Running a priori methods online (periodically) could conceivably capture dynamic competency changes. However, such assessments can waste computational resources if competency has, in fact, not changed or may be too expensive for certain kinds of decision-making agents ( Acharya et al, 2022 ; Conlon et al, 2022a ; Gautam et al, 2022 ).…”

“…Another method of in situ self-assessment involves monitoring features of the agent’s current state. For example, Gautam et al (2022) developed a method to monitor deviations from design assumptions, while Ramesh et al (2022) used the “vitals” of a robot to monitor its health during task execution. Both methods provide a valuable instantaneous snapshot of the agent’s state at a given time, which can indicate performance degradation online; however, neither predicts higher-level task competency (for example, does the degradation actually impact the task outcome?).…”