Adaptive Learning for Whole-body Navigation and Manipulation with Shipboard Robots
Abstract
Adaptive Learning for Whole-body Navigation and Manipulation by Shipboard RobotsResearch Problem and Objectives: We envision autonomous multi-limbed robots that traverse tight spaces to perform shipboard maintenance and service tasks. To be effective in these highlydynamic scenarios, robots need the ability to adapt their navigation and manipulation algorithms to the task-at-hand. For example, a robot tasked with replacing a cable may need to reach in to atight space to plug in the cable with one arm while balancing itself with its remaining limbs. In order to perform such tasks the robot would use planning and control algorithms to generate its motion, however these algorithms require models of the dynamics of the objects being manipulated as well the dynamics of contact between the robots limbs and surfaces. Unfortunately thisknowledge will not be available to the robot a priori, and thus existing methods that rely on analytical modeling or large datasets to construct dynamics models cannot be applied. To overcome this challenge, we propose to develop online learning, control, and planning methods that are able to use prior knowledge and adapt that knowledge to the new task-at-hand with very limited datagathered online.Technical Approaches: This is a challenging problem, as we cannot hope to find globally-accurate dynamics models with limited data. Instead we propose to develop methods that mix controllers based on simplified models of objects (trained with offline data) with local controllers derived from the data we gather online in a model-predictive control fashion. However, control alone will notbe sufficient to perform tasks such as traversing a staircase or servicing wiring; the robot must also be able to construct long-term plans. To do this we will build on our newly-defined Blindfolded Travelers Problem (BTP), a novel graph search problem where edgesorm a task given uncertainty estimates on the success ofactions. This formulation will allow us to intelligently trade off success probability with efficiency while performing a fast search. We will test and verify the efficacy of our methods on three Navyrelevant tasks: 1) Replacing cabling in tight spaces; 2) Traversing a tight cluttered hallway where removing obstructions is necessary; and 3) Gathering and hauling rope.Anticipated Outcome: The expected outcome of this research is a framework for whole-body navigation and manipulation in novel scenarios that allows a robot to adapt prior knowledge to novel tasks-at-hand. Given prior knowledge of dynamics in a similar situation, the robot will be able to adapt that knowledge for use in control and high-level planning while taking into account the uncertainty in its adapted dynamics models. Impact on DoD Capabilities: Regular maintenance of equipment is essential to reduce the need to replace costly components and to maintain military readiness. We expect this research to provide a significant step toward autonomous shipboard robots that can perform many routine maintenance tasks that are time consuming, tedious, and/or costly to perform for Navy personnel. The proposed algorithms, when deployed on shipboard robots, could also make performance of maintenance tasks more consistent and reliable while allowing Navy personnel to focus on their primary duties.The proposed work can also have impact on deck operations by enabling robots to haul and organize rope and lines. Finally, shipboard services involving deformable objects, e.g. cooking and laundry, could be done more efficiently and quickly using shipboard robots endowed with the proposed manipulation and navigation capabilities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 02, 2021
- Source ID
- N000142112118
Entities
People
- Dmitry Berenson
Organizations
- Board of Regents of the University of Michigan
- Office of Naval Research
- United States Navy