O2A: One-Shot Observational Learning with Action Vectors

We present O₂A, a novel method for learning to perform robotic manipulation tasks from a single (one-shot) third-person demonstration video. To our knowledge, it is the first time this has been done for a single demonstration. The key novelty lies in pre-training a feature extractor for creating a perceptual representation for actions that we call “action vectors”. The action vectors are extracted using a 3D-CNN model pre-trained as an action classifier on a generic action dataset. The distance between the action vectors from the observed third-person demonstration and trial robot executions is used as a reward for reinforcement learning of the demonstrated task. We report on experiments in simulation and on a real robot, with changes in viewpoint of observation, properties of the objects involved, scene background and morphology of the manipulator between the demonstration and the learning domains. O₂A outperforms baseline approaches under different domain shifts and has comparable performance with an Oracle (that uses an ideal reward function). Videos of the results, including demonstrations, can be found in our: project-website.

Introducing the NEMO-Lowlands iconic gesture dataset, collected through a gameful human-robot interaction

This paper describes a novel dataset of iconic gestures, together with a publicly available robot-based elicitation method to record these gestures, which consists of playing a game of charades with a humanoid robot. The game was deployed at a science museum (NEMO) and a large popular music festival (Lowlands) in the Netherlands. This resulted in recordings of 428 participants, both adults and children, performing 3715 silent iconic gestures for 35 different objects in a naturalistic setting. Our dataset adds to existing collections of iconic gesture recordings in two important ways. First, participants were free to choose how they represented the broad concepts using gestures, and they were asked to perform a second attempt if the robot did not recognize their gesture the first time. This provides insight into potential repair strategies that might be used. Second, by making the interactive game available we enable other researchers to collect additional recordings, for different concepts, and in diverse cultures or contexts. This can be done in a consistent manner because a robot is used as a confederate in the elicitation procedure, which ensures that every data collection session plays out in the same way. The current dataset can be used for research into human gesturing behavior, and as input for the gesture recognition and production capabilities of robots and virtual agents.

Human Motion Understanding for Selecting Action Timing in Collaborative Human-Robot Interaction

In the industry of the future, so as in healthcare and at home, robots will be a familiar presence. Since they will be working closely with human operators not always properly trained for human-machine interaction tasks, robots will need the ability of automatically adapting to changes in the task to be performed or to cope with variations in how the human partner completes the task. The goal of this work is to make a further step toward endowing robot with such capability. To this purpose, we focus on the identification of relevant time instants in an observed action, called dynamic instants, informative on the partner's movement timing, and marking instants where an action starts or ends, or changes to another action. The time instants are temporal locations where the motion can be ideally segmented, providing a set of primitives that can be used to build a temporal signature of the action and finally support the understanding of the dynamics and coordination in time. We validate our approach in two contexts, considering first a situation in which the human partner can perform multiple different activities, and then moving to settings where an action is already recognized and shows a certain degree of periodicity. In the two contexts we address different challenges. In the first one, working in batch on a dataset collecting videos of a variety of cooking activities, we investigate whether the action signature we compute could facilitate the understanding of which type of action is occurring in front of the observer, with tolerance to viewpoint changes. In the second context, we evaluate online on the robot iCub the capability of the action signature in providing hints to establish an actual temporal coordination during the interaction with human participants. In both cases, we show promising results that speak in favor of the potentiality of our approach.

Learning Semantics of Gestural Instructions for Human-Robot Collaboration

Designed to work safely alongside humans, collaborative robots need to be capable partners in human-robot teams. Besides having key capabilities like detecting gestures, recognizing objects, grasping them, and handing them over, these robots need to seamlessly adapt their behavior for efficient human-robot collaboration. In this context we present the fast, supervised Proactive Incremental Learning (PIL) framework for learning associations between human hand gestures and the intended robotic manipulation actions. With the proactive aspect, the robot is competent to predict the human's intent and perform an action without waiting for an instruction. The incremental aspect enables the robot to learn associations on the fly while performing a task. It is a probabilistic, statistically-driven approach. As a proof of concept, we focus on a table assembly task where the robot assists its human partner. We investigate how the accuracy of gesture detection affects the number of interactions required to complete the task. We also conducted a human-robot interaction study with non-roboticist users comparing a proactive with a reactive robot that waits for instructions.