We consider the problem of detecting robotic grasps in an RGB-D view of a scene containing objects. In this work, we apply a deep learning approach to solve this problem, which avoids time-consuming hand-design of features. This presents two main challenges. First, we need to evaluate a huge number of candidate grasps. In order to make detection fast and robust, we present a two-step cascaded system with two deep networks, where the top detections from the first are re-evaluated by the second. The first network has fewer features, is faster to run, and can effectively prune out unlikely candidate grasps. The second, with more features, is slower but has to run only on the top few detections. Second, we need to handle multimodal inputs effectively, for which we present a method that applies structured regularization on the weights based on multimodal group regularization. We show that our method improves performance on an RGBD robotic grasping dataset, and can be used to successfully execute grasps on two different robotic platforms.

Deep Learning for Detecting Robotic Grasps

As robots and other intelligent agents move from simple environments and problems to more complex, unstructured settings, manually programming their behavior has become increasingly challenging and expensive. Often, it is easier for a teacher to demonstrate a desired behavior rather than attempt to manually engineer it. This process of learning from demonstrations, and the study of algorithms to do so, is called imitation learning. This work provides an introduction to imitation learning. It covers the underlying assumptions, approaches, and how they relate; the rich set of algorithms developed to tackle the problem; and advice on effective tools and implementation. We intend this paper to serve two audiences. First, we want to familiarize machine learning experts with the challenges of imitation learning, particularly those arising in robotics, and the interesting theoretical and practical distinctions between it and more familiar frameworks like statistical supervised learning theory and reinforcement learning. Second, we want to give roboticists and experts in applied artificial intelligence a broader appreciation for the frameworks and tools available for imitation learning. We pay particular attention to the intimate connection between imitation learning approaches and those of structured prediction Daume III et al. [2009]. To structure this discussion, we categorize imitation learning techniques based on the following key criteria which drive algorithmic decisions:

1) The structure of the policy space. Is the learned policy a time-index trajectory (trajectory learning), a mapping from observations to actions (so called behavioral cloning [Bain and Sammut, 1996]), or the result of a complex optimization or planning problem at each execution as is common in inverse optimal control methods [Kalman, 1964, Moylan and Anderson, 1973].

2) The information available during training and testing. In particular, is the learning algorithm privy to the full state that the teacher possess? Is the learner able to interact with the teacher and gather corrections or more data? Does the learner have a (typically a priori) model of the system with which it interacts? Does the learner have access to the reward (cost) function that the teacher is attempting to optimize?

3) The notion of success. Different algorithmic approaches provide varying guarantees on the resulting learned behavior. These guarantees range from weaker (e.g., measuring disagreement with the agent’s decision) to stronger (e.g., providing guarantees on the performance of the learner with respect to a true cost function, either known or unknown). We organize our work by paying particular attention to distinction (1): dividing imitation learning into directly replicating desired behavior (sometimes called behavioral cloning) and learning the hidden objectives of the desired behavior from demonstrations (called inverse optimal control or inverse reinforcement learning [Russell, 1998]). In the latter case, behavior arises as the result of an optimization problem solved for each new instance that the learner faces. In addition to method analysis, we discuss the design decisions a practitioner must make when selecting an imitation learning approach. Moreover, application examples—such as robots that play table tennis [Kober and Peters, 2009], programs that play the game of Go [Silver et al., 2016], and systems that understand natural language [Wen et al., 2015]— illustrate the properties and motivations behind different forms of imitation learning. We conclude by presenting a set of open questions and point towards possible future research directions for machine learning.

/pdf/an-algorithmic-perspective-on-imitation-learning-n55kqg2cp5.pdf

An Algorithmic Perspective on Imitation Learning

BACKGROUND Humans have a fantastic ability to manipulate objects of various shapes, sizes, and materials and can control the objects’ position in confined spaces with the advanced dexterity capabilities of our hands. Building machines inspired by human hands, with the functionality to autonomously pick up and manipulate objects, has always been an essential component of robotics. The first robot manipulators date back to the 1960s and are some of the first robotic devices ever constructed. In these early days, robotic manipulation consisted of carefully prescribed movement sequences that a robot would execute with no ability to adapt to a changing environment. As time passed, robots gradually gained the ability to automatically generate movement sequences, drawing on artificial intelligence and automated reasoning. Robots would stack boxes according to size, weight, and so forth, extending beyond geometric reasoning. This task also required robots to handle errors and uncertainty in sensing at run time, given that the slightest imprecision in the position and orientation of stacked boxes might cause the entire tower to topple. Methods from control theory also became instrumental for enabling robots to comply with the environment’s natural uncertainty by empowering them to adapt exerted forces upon contact. The ability to stably vary forces upon contact expanded robots’ manipulation repertoire to more-complex tasks, such as inserting pegs in holes or hammering. However, none of these actions truly demonstrated fine or in-hand manipulation capabilities, and they were commonly performed using simple two-fingered grippers. To enable multipurpose fine manipulation, roboticists focused their efforts on designing humanlike hands capable of using tools. Wielding a tool in-hand became a problem of its own, and a variety of advanced algorithms were developed to facilitate stable holding of objects and provide optimality guarantees. Because optimality was difficult to achieve in a stochastic environment, from the 1990s onward researchers aimed to increase the robustness of object manipulation at all levels. These efforts initiated the design of sensors and hardware for improved control of hand–object contacts. Studies that followed were focused on robust perception for coping with object occlusion and noisy measurements, as well as on adaptive control approaches to infer an object’s physical properties, so as to handle objects whose properties are unknown or change as a result of manipulation. ADVANCES Roboticists are still working to develop robots capable of sorting and packaging objects, chopping vegetables, and folding clothes in unstructured and dynamic environments. Robots used for modern manufacturing have accomplished some of these tasks in structured settings that still require fences between the robots and human operators to ensure safety. Ideally, robots should be able to work side by side with humans, offering their strength to carry heavy loads while presenting no danger. Over the past decade, robots have gained new levels of dexterity. This enhancement is due to breakthroughs in mechanics with sensors for perceiving touch along a robot’s body and new mechanics for soft actuation to offer natural compliance. Most notably, this development leverages the immense progress in machine learning to encapsulate models of uncertainty and support further advances in adaptive and robust control. Learning to manipulate in real-world settings is costly in terms of both time and hardware. To further elaborate on data-driven methods but avoid generating examples with real, physical systems, many researchers use simulation environments. Still, grasping and dexterous manipulation require a level of reality that existing simulators are not yet able to deliver—for example, in the case of modeling contacts for soft and deformable objects. Two roads are hence pursued: The first draws inspiration from the way humans acquire interaction skills and prompts robots to learn skills from observing humans performing complex manipulation. This allows robots to acquire manipulation capabilities in only a few trials. However, generalizing the acquired knowledge to apply to actions that differ from those previously demonstrated remains difficult. The second road constructs databases of real object manipulation, with the goal to better inform the simulators and generate examples that are as realistic as possible. Yet achieving realistic simulation of friction, material deformation, and other physical properties may not be possible anytime soon, and real experimental evaluation will be unavoidable for learning to manipulate highly deformable objects. OUTLOOK Despite many years of software and hardware development, achieving dexterous manipulation capabilities in robots remains an open problem—albeit an interesting one, given that it necessitates improved understanding of human grasping and manipulation techniques. We build robots to automate tasks but also to provide tools for humans to easily perform repetitive and dangerous tasks while avoiding harm. Achieving robust and flexible collaboration between humans and robots is hence the next major challenge. Fences that currently separate humans from robots will gradually disappear, and robots will start manipulating objects jointly with humans. To achieve this objective, robots must become smooth and trustable partners that interpret humans’ intentions and respond accordingly. Furthermore, robots must acquire a better understanding of how humans interact and must attain real-time adaptation capabilities. There is also a need to develop robots that are safe by design, with an emphasis on soft and lightweight structures as well as control and planning methodologies based on multisensory feedback.

Trends and challenges in robot manipulation

Robotics and autonomous systems

IEEE International Conference on Robotics and Automation

We address the difficult problem of catching in-flight objects with uneven shapes This requires the solution of three complex problems: accurate prediction of the trajectory of fastmoving objects, predicting the feasible catching configuration, and planning the arm motion, and all within milliseconds We follow a programming-by-demonstration approach in order to learn, from throwing examples, models of the object dynamics and arm movement We propose a new methodology to find a feasible catching configuration in a probabilistic manner We use the dynamical systems approach to encode motion from several demonstrations This enables a rapid and reactive adaptation of the arm motion in the presence of sensor uncertainty We validate the approach in simulation with the iCub humanoid robot and in real-world experiments with the KUKA LWR 4+ (7-degree-of-freedom arm robot) to catch a hammer, a tennis racket, an empty bottle, a partially filled bottle, and a cardboard box

Catching Objects in Flight

Coupled dynamical system based arm-hand grasping model for learning fast adaptation strategies

A direct variational method for planning monotonically optimal paths for redundant manipulators in constrained workspaces

Background The ability to follow one another's gaze plays an important role in our social cognition; especially when we synchronously perform tasks together. We investigate how gaze cues can improve performance in a simple coordination task (i.e., the mirror game), whereby two players mirror each other's hand motions. In this game, each player is either a leader or follower. To study the effect of gaze in a systematic manner, the leader's role is played by a robotic avatar. We contrast two conditions, in which the avatar provides or not explicit gaze cues that indicate the next location of its hand. Specifically, we investigated (a) whether participants are able to exploit these gaze cues to improve their coordination, (b) how gaze cues affect action prediction and temporal coordination, and (c) whether introducing active gaze behavior for avatars makes them more realistic and human-like (from the user point of view). Methodology/Principal Findings 43 subjects participated in 8 trials of the mirror game. Each subject performed the game in the two conditions (with and without gaze cues). In this within-subject study, the order of the conditions was randomized across participants, and subjective assessment of the avatar's realism was assessed by administering a post-hoc questionnaire. When gaze cues were provided, a quantitative assessment of synchrony between participants and the avatar revealed a significant improvement in subject reaction-time (RT). This confirms our hypothesis that gaze cues improve the follower's ability to predict the avatar's action. An analysis of the pattern of frequency across the two players' hand movements reveals that the gaze cues improve the overall temporal coordination across the two players. Finally, analysis of the subjective evaluations from the questionnaires reveals that, in the presence of gaze cues, participants found it not only more human-like/realistic, but also easier to interact with the avatar. Conclusion/Significance This work confirms that people can exploit gaze cues to predict another person's movements and to better coordinate their motions with their partners, even when the partner is a computer-animated avatar. Moreover, this study contributes further evidence that implementing biological features, here task-relevant gaze cues, enable the humanoid robotic avatar to appear more human-like, and thus increase the user's sense of affiliation.

/pdf/role-of-gaze-cues-in-interpersonal-motor-coordination-15piy8d6yv.pdf

Role of Gaze Cues in Interpersonal Motor Coordination: Towards Higher Affiliation in Human-Robot Interaction.

We integrate software components that allow efficient and successful grasping of kitchenware objects. The contributed components include: The object pose detector, the gripper reaching motion and the grasp hypothesis selection. The object pose detector of Drost et. al. [10] is improved, considering rotationally symmetric objects. The reaching motion execution combines two independent dynamical systems: The approach direction system and its tangent space [21]. The coupling provides a robust reaching component that copes with several gripper configurations. The grasp hypothesis selection filters the object poses by considering the table orientation.

/pdf/reaching-and-grasping-kitchenware-objects-4f5fh02nr6.pdf

Ashwini Shukla

Papers

Catching Objects in Flight

Coupled dynamical system based arm-hand grasping model for learning fast adaptation strategies

A direct variational method for planning monotonically optimal paths for redundant manipulators in constrained workspaces

Role of Gaze Cues in Interpersonal Motor Coordination: Towards Higher Affiliation in Human-Robot Interaction.

Reaching and grasping kitchenware objects