Generative Models for Tool Use from Robot Play
Abstract
The proposed work will develop methods for robots to `play , engaging in behaviors that are discovery oriented rather than practice, competition, or task oriented. Tool use will be a particular focus, developing algorithmic approaches for a robot to create a library of robot/tool and tool/object interactions to enable planning for tool use. Our model of this discovery process is that the robot must discover its own interactions with a tool and then discover interactions between the tool and objects in the robot s environment, but without advance knowledge of an intended task and without advance knowledge of the dynamics of the tool. These interactions will be modeled using generative methods from machine learning, but the robot s behavior will not be randomized for the purpose of learning. Instead, the robot will synthesize its actions with the goal of improving its understanding of the physical interactions themselves, without any specified task. The data collected in this `tinkering phase of exploring without intent will form the basis for segmenting interactions into a library of causal models that can be sequenced to achieve a specific task. The goal is to formalize physical activities we often take for granted in animal behavior, including the ability to `tinker with objects and within the environment without specific, task-oriented intent, and to `play with the resulting capabilities. These activities enable new modalities of interaction that may facilitate later task-specific goals. One of the key aspects of the work is that we assume that imagination (i.e., the ability to foresee many plausible futures through simulation) is explicitly of limited utility (perhaps because the world cannot be simulated effectively, or because the statistical properties of the simulations are inadequate for learning). Instead, we assume that a robot is using what would otherwise be its extensive down time to build up capabilities in a very long, but nevertheless single-shot, epoch of experience and learning. In this setting, simulation explicitly takes on an inferior role to physical experience. Our technical approach will rely on highly automated exploration techniques, representations from machine learning suitable for use in combination with exploration in real-time, one-shot learning applications, and tools from nonsmooth optimization in control and design. Ergodic measures and associated ergodic control algorithms play a central role in automating the dynamically-evolving coverage of information landscapes needed for active learning. Conditional variational autoencoders (CVAEs) will be our model of machine learning. In the context of these two mathematical and algorithmic choices, we will determine mechanisms for safe exploration, guarantees on convergence, and design automation using nonsmooth optimization techniques. Moreover, we will use simulation and hardware experiments to ground the theoretical and algorithmic development. Specifically, the work will focus on unconventional soft body tools like bendable wire. Unlike traditional tools (e.g., a hammer) that have a particular, intended use (hammering a nail), wire is flexible and can be shaped to many different purposes depending on the task. For example, a wire can be used to hook around something that is out of reach or to form a cage around something too dangerous (e.g., hot) for the robot to handle. The experimental goal is to demonstrate in simulation and in hardware that a robot with lots of time available for physical learning can generate representations of creative tool use. After developing algorithms for the bendable wire tool, we will apply the same formulations to more traditional tools (e.g., an adjustable wrench) to discover both intended use (e.g., gripping) and unintended use (e.g., hammering a nail, puncturing a balloon, bending a wire).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 12, 2022
- Source ID
- W911NF2210286
Entities
People
- Todd D Murphey
Organizations
- Army Contracting Command
- Northwestern University
- United States Army