Risk-Aware, Human-Cooperative Planning for Autonomous Systems
Technical point of contact: ?? , Office of Naval Research
Period of activity: 2014-2017
Collaborators: Masahiro Ono (Lead, JPL), Behcet Acikmese (Univ of Washington), Missy Cummings (Duke)
Overview of the Project
The ability to manage risk is an indispensable part of human and machine intelligence when performing complex tasks, ranging from military operations to space exploration. Although machine intelligence plays increasingly significant roles, in most cases humans are still solely responsible for predicting and coping with risks, while robots simply execute a given plan without explicit awareness of risk. This project aims to revise the relationship between human and robot to a cooperative partnership, where both parties share the responsibility of managing risks. This paradigm shift will not just mitigate the cognitive workload of human operators but also make a human-robot system significantly safer and more reliable because robots can quickly respond to contingencies without waiting for the instructions from human operators and humans and robots can cover the weaknesses of each other. We call this new concept of machine intelligence as Risk-aware Human-cooperative Autonomy (RHA).
In the proposed RHA concept, the instructions from human operators are not a detailed command sequence but high-level goals as well as bounds on risks. RHA is responsible of optimizing the actions of robots to achieve the goals within the risk bounds, while flexibly responding to contingencies. It also continuously informs humans of risk assessments, and accepts changes in goals and risk bounds if necessary. We seek answers to answer the following two fundamental questions:
- What is the most productive way for human and autonomy to cooperate in a risky situation? How humans can be best informed of risks? What is the natural way to instruct autonomy about risks? How human and autonomy should divide the responsibility of handling risks?
- How to realize a risk-aware autonomy that is reliable, tractable, and capable of accepting high-level instructions on goals and risks? How to scale it to real-world problems without sacrificing fidelity and correctness?