Abstract
To endow robots with the flexibility to perform a wide range of tasks in diverse and complex environments, learning their controller from experience data is a promising approach. In particular, some recent meta-learning methods are shown to solve novel tasks by leveraging their experience of performing other tasks during training. Although studies around meta-learning of robot control have worked on improving the performance, the safety issue has not been fully explored, which is also an important consideration in the deployment. In this paper, we firstly relate uncertainty on task inference with the safety in meta-learning of visual imitation, and then propose a novel framework for estimating the task uncertainty through probabilistic inference in the task-embedding space, called PETNet. We validate PETNet with a manipulation task with a simulated robot arm in terms of the task performance and uncertainty evaluation on task inference. Following the standard benchmark procedure in meta-imitation learning, we show PETNet can achieve the same or higher level of performance (success rate of novel tasks at meta-test time) as previous methods. In addition, by testing PETNet with semantically inappropriate or synthesized out-of-distribution demonstrations, PETNet shows the ability to capture the uncertainty about the tasks inherent in the given demonstrations, which allows the robot to identify situations where the controller might not perform properly. These results illustrate our proposal takes a significant step forward to the safe deployment of robot learning systems into diverse tasks and environments.
Highlights
The development of generalist robots remains a key challenge in robotics
The performance degrades if we use demonstrations with high task uncertainty, and increasing the number of demonstrations for adaptation can help reduce the task uncertainty. These results show that Probabilistic Embedding over Task-space Network (PETNet) can contribute to the safety of meta-imitation learning for controllers by preventing robots from behaving unexpectedly when the demonstrations have too much uncertainty for robots to identify the task
We first pointed out the significance of evaluating task uncertainty in meta-learning for robot control to ensure safe deployment
Summary
The development of generalist robots remains a key challenge in robotics. These robots are expected to perform a wide range of tasks in diverse environments, for instance, automation of household chores or operations in retail stores. The performance degrades if we use demonstrations with high task uncertainty, and increasing the number of demonstrations for adaptation can help reduce the task uncertainty These results show that PETNet can contribute to the safety of meta-imitation learning for controllers by preventing robots from behaving unexpectedly when the demonstrations have too much uncertainty for robots to identify the task. Key contributions of this paper are (1) proposing a framework for leveraging uncertainty of task indicated with demonstrations for realizing the safety of robot under various tasks in meta-imitation learning and (2) with the simple implementation with probabilistic inference on task embedding space, this approach can capture the task uncertainty, leading to identify demonstrations with which the model would not perform properly
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
