Machine Learning for Robot Grasping and Manipulation

Robotics as a technology has an incredible potential for improving our everyday lives. Robots could perform household chores, such as cleaning, cooking, and gardening, in order to give us more time for other pursuits. Robots could also be used to perform tasks in hazardous environments, such as turning off a valve in an emergency or safely sorting our more dangerous trash. However, all of these applications would require the robot to perform manipulation tasks with various objects. Today's robots are used primarily for performing specialized tasks in controlled scenarios, such as manufacturing. The robots that are used in today's applications are typically designed for a single purpose and they have been preprogrammed with all of the necessary task information. In contrast, a robot working in a more general environment will often be confronted with new objects and scenarios. Therefore, in order to reach their full potential as autonomous physical agents, robots must be capable of learning versatile manipulation skills for different objects and situations. Hence, we have worked on a variety of manipulation skills to improve those capabilities of robots, and the results have lead to several new approaches, which are presented in this thesis Learning manipulation skills is, however, an open problem with many challenges that still need to be overcome. The first challenge is to acquire and improve manipulation skills with little to no human supervision. Rather than being preprogrammed, the robot should be able to learn from human demonstrations and through physical interactions with objects. Learning to improve skills through trial and error learning is a particularly important ability for an autonomous robot, as it allows the robot to handle new situations. This ability also removes the burden from the human demonstrator to teach a skill perfectly, as a robot is allowed to make mistakes if it can learn from them. In order to address this challenge, we present a continuum-armed bandits approach for learning to grasp objects. The robot learns to predict the performances of different grasps, as well as how certain it is of this prediction, and selects grasps accordingly. As the robot tries more grasps, its predictions become more accurate, and its grasps improve accordingly. A robot can master a manipulation skill by learning from different objects in various scenarios. Another fundamental challenge is therefore to efficiently generalize manipulations between different scenarios. Rather than relearning from scratch, the robot should find similarities between the current situation and previous scenarios in order to reuse manipulation skills and task information. For example, the robot can learn to adapt manipulation skills to new objects by finding similarities between them and known objects. However, only some similarities between objects will be relevant for a given manipulation. The robot must therefore also learn which similarities are important for adapting the manipulation skill. We present two object representations for generalizing between different situations. Contacts between objects are important for many manipulations, but it is difficult to define general features for representing sets of contacts. Instead, we define a kernel function for comparing contact distributions, which allows the robot to use kernel methods for learning manipulations. The second approach is to use warped parameters to define more abstract features, such as areas and volumes. These features are defined as functions of known object models. The robot can compute these parameters for novel objects by warping the shape of the known object to match the unknown object. Learning about objects also requires the robot to reconcile information from multiple sensor modalities, including touch, hearing, and vision. While some object properties will only be observed by specific sensor modalities, other object properties can be determined from multiple sensor modalities. For example, while color can only be determined by vision, the shape of an object can be observed using vision or touch. The robot should use information from all of its senses in order to quickly learn about objects. We explain how the robot can learn low-dimensional representations of tactile data by incorporating cues from vision data. As touching an object usually occludes the surface, the proposed method was designed to work with weak pairings between the data in the two sensor modalities. The robot can also learn more efficiently if it reuses skills between different tasks. Rather than relearn a skill for each new task, the robot should learn manipulation skills that can be reused for multiple tasks. For an autonomous robot, this would require the robot to divide tasks into smaller steps. Dividing tasks into smaller parts makes it easier to learn the corresponding skills. If a step is a part of many tasks, then the robot will have more opportunities to practice the associated skill, and more tasks will benefit from the resulting performance improvement. In order to learn a set of useful subtasks, we propose a probabilistic model for dividing manipulations into phases. This model captures the conditions for transitioning between different phases, which represent subgoals and constraints of the overall tasks. The robot can use the model together with model-based reinforcement learning in order to learn skills for moving between phases. When confronted with a new task, the robot will have to select a suitable sequence of skills to execute. The robot must therefore also learn to select which manipulation to execute in the current scenario. Selecting sequences of motor primitives is difficult, as the robot must take into consideration the current task, state, and future actions when selecting the next motor skill to execute. We therefore present a value function method for selecting skills in an optimal manner. The robot learns the value function for the continuous state space using a flexible non-parametric model-based approach. Learning manipulation skills also poses certain challenges for learning methods. The robot will not have thousands of samples when learning a new manipulation skill, and must instead actively collect new samples or use data from similar scenarios. The learning methods presented in this thesis are, therefore, designed to work with relatively small amounts of data, and can generally be used during the learning process. Manipulation tasks also present a spectrum of different problem types. Hence, we present supervised, unsupervised, and reinforcement learning approaches in order to address the diverse challenges of learning manipulations skills.