Safe Task Execution by Autonomous Humanoid Robots
(TÜBİTAK 1001, 01/09/2015 - 01/07/2018)
A robot should detect failures or anomalies and recover from them for safe execution to prevent potential damages to its environment or objects in interest. In this project, we investigate failure detection, isolation/identification and recovery methods for safe task execution by autonomous humanoid robots. The main motivation behind our work is developing methods that enable robots’ safe use in everyday application scenarios. Although this is a well-studied research topic for industrial robots in structured and engineered environments, for which certain standards and regulations exist, there are still several open research questions for safety in task execution by autonomous robots working in unstructured environments. We address these questions and propose a safe task execution system for solving them.
 
Our system will include a perception pipeline to efficiently process data from different sensor modalities and extract useful information to monitor task execution. In particular, useful signal features from audio-visual and force/tactile sensor data will be considered, and predicates on world facts will be maintained to represent failure contexts corresponding to both external and internal states of the robot. Probabilistic sensor fusion methods will be applied for this purpose. Furthermore, adaptive active sensing methods will be investigated to determine the effects of each sensory data component on certain conditions and to apply a computationally efficient active selection strategy for sensing.
 
The main contribution of the system lies on the targeted failure detection and isolation/identification methods. A Metric Temporal Logic (MTL) rule-based failure detection method will be applied. Continually monitored predicates and relations will take place in these rules. The parameters used in these rules will be learned by reinforcement learning methods resulting in adaptive rules for different conditions.
 
For failure isolation both supervised learning methods and temporal probabilistic models will be investigated. A Hierarchical Hidden Markov Model (HHMM)-based failure isolation method will explain failure cases with relevant reasons. Failure models will be represented as distinct HHMMs. These models will be updated during the execution of actions in a plan while the environment is continually monitored. Object recognition, object localization, execution parameter/type, hardware, collision failures and external events will be modeled by this system. Eventually, the system will classify the situation as a success, a fail-safe or a fail-unsafe state. After isolating a failure, the robot will recover from the unexpected situation by either changing its execution parameters or re-planning. This type of classification is performed by using such a broad set of sensors for the first time with this project.
 
Failure detection and recovery in unstructured environments is essential for safe task execution in everyday scenarios by autonomous robots. The results from this study are expected to have broad impact on autonomous robots and human-robot interactions research featuring safety and reliability in both home and work environments.