Combined Task and Action Learning from Human Demonstrations for Mobile Manipulation Applications

Abstract

Learning from demonstrations is a promising paradigm for transferring knowledge to robots. However, learning mobile manipulation tasks directly from a human teacher is a complex problem as it requires learning models of both the overall task goal and of the underlying actions. Additionally, learning from a small number of demonstrations often introduces ambiguity with respect to the intention of the teacher, making it challenging to commit to one model for generalizing the task to new settings. In this paper, we present an approach to learning flexible mobile manipulation action models and task goal representations from teacher demonstrations. Our action models enable the robot to consider different likely outcomes of each action and to generate feasible trajectories for achieving them. Accordingly, we leverage a probabilistic framework based on Monte Carlo tree search to compute sequences of feasible actions imitating the teacher intention in new settings without requiring the teacher to specify an explicit goal state. We demonstrate the effectiveness of our approach in complex tasks carried out in real-world settings.

Task-Oriented Motion Mapping on Robots of Various Configuration Using Body Role Division

“…Welschehold et al approached the problem by mapping the human torso movement for a single arm task [8]. They have further extended their work to learn individual actions and disambiguate overall task goals in a mobile setting from a small number of demonstrations [9].…”

Error-Aware Imitation Learning from Teleoperation Data for Mobile Manipulation

“…Most related to our work, several works have leveraged IL for MM tasks [5,6,7,8]. [5] presented a web-based tool for crowdsourcing a large scale dataset of MM tasks, and used it in combination with motion planning for execution on the robot.…”

“…[5] presented a web-based tool for crowdsourcing a large scale dataset of MM tasks, and used it in combination with motion planning for execution on the robot. [6] and [7] collected RGBD observations of humans performing tasks such as door opening and tabletop object manipulation, and used hypergraph optimization and a search procedure respectively to adapt these trajectories to be executable by a robot. Lastly, [8] collected VR demonstrations of pick and place actions and extracted a sequence of symbolic actions and action parametrizations to adopt them for use on a robot.…”

“…Unfortunately, the vast state space in MM has inhibited data collection from humans due to the difficulty of annotating substantial portions of the state space with good actions. Prior work has avoided this issue by instrumenting the environment to ease the state coverage burden [5,6,7,8]. However, even if there were a way to collect full human demonstrations without instrumentation, learning from such datasets would also pose a challenge due to covariate shift [9], where trained policies visit states not covered by the demonstrations -an outcome that is more likely in MM.…”