Learning-Based Robot Control in Data-Limited Environments

Abstract

Manufacturing plays a key role in the economy, and robotics plays a key role in manufacturing. Developments in both areas are soaring, driven in part by growing caches of data and increasingly intelligent algorithms that capitalize on data. However, there remains a misalignment between the pace of robotic advancement and the ability of industry to fully benefit. Practical manufacturing systems often do not satisfy the data requirements of advanced robot learning algorithms, either in the form of rich offline data or high-resolution sensing. Small- and medium-sized enterprises (SMEs), in particular, are short on data and sensitive to the cost of additional instrumentation. This dissertation focuses on how industrial robots can be endowed with the intelligence to autonomously learn to complete common industrial tasks, subject the limited data available in SMEs’ manufacturing systems. Three levels of data availability are considered.

At the lowest level of data availability are robots that operate in unmodeled environments without sensing and with only a small historical dataset of prior robot programs. A framework is developed to leverage this dataset to enable automatic generation of programs for new tasks. From the historical dataset, the framework learns (1) how to interact with objects, (2) how to move the robot, and (3) where the robot can move safely. The framework can then generate new open-loop robot programs, along with an estimate of performance. The framework is shown to scale from a small dataset of programs to a much larger number of feasible tasks.

At the next level of data availability are robots that perform repetitive tasks and have temporary access to a mobile sensor. Given a new task in an unmodeled environment, the robot and sensor must cooperate the complete the task, while minimizing the amount of time that the robot requires the sensing resource. During the cooperation, the sensor places itself for optimal data collection while the robot finds a controller to complete the task safely with the available data. A hierarchical control scheme is proposed, involving a differentiable controller and sensor model. The control scheme allows both the controller parameters and sensor pose to be optimized simultaneously and efficiently, even when the sensor and controller are poorly initialized.

At the final level of data availability are robots that perform unstructured tasks requiring visual feedback. In particular, wire harness installation, a common unstructured task in automotive manufacturing yet to be automated, is considered. Since task parameters may change between iterations, learning must occur within a single task iteration. Despite the apparent complexity, an approximate control-oriented parametric model of the harness dynamics is derived, which adapts to specific harnesses in situ without prior data. A model predictive controller is developed to perform the manipulation using the adaptive model, which is experimentally validated in simulation and reality.

Across the three regimes of data availability, the contributions of this dissertation maximize the utility and flexibility of robots in manufacturing. Robotic tasks that previously required a combination of human labor, rich datasets, and high-quality sensing can now be automated in far more accessible ways. As a result, manufacturers both large and small can effectively scale up without significant capital investment, leading to the economic growth potential afforded by a flexible and agile manufacturing industry.