Shared Autonomy via Robust Task Planning and Argumentation (SHARPA)
Lead Research Organisation:
Royal Holloway University of London
Department Name: Computer Science
Abstract
The overarching objective of this project is to endow autonomous systems with advanced decision-making capabilities and collaboration skills. We aim to build artificial systems that, in real-world environments, are capable of reasoning about high-level goals specified by human operators and formulating, in collaboration with them, a course of actions to successfully achieve such goals. Strategic reasoning and fluid teaming are fundamental skills of cognitive systems: they are needed in a variety of situations, from day-to-day tasks such as assisting humans in household chores, to extreme missions, such as space exploration. The techniques that we propose are general and can be used to support both robotic systems and software agents. We choose disaster response operations where unmanned aerial vehicles (UAVs) assist emergency responders as our demonstration arena. In this domain, in fact, it is crucial for the UAVs to think strategically to pursue goals efficiently and to act in concert with the human operators who are ultimately in charge of critical decisions.
The primary objective of the project is broken down into two strands. The first is to equip artificial artefacts that operate in real-world settings with the ability to reason about themselves and the world around them to determine plans for achieving high-level goals efficiently and robustly. Planning is a key component of intelligence and one of the most traditional fields of artificial intelligence (AI). Planning has achieved impressive results in idealised settings where the world is deterministic, and actions are instantaneous. However, planning in real-world environments in which temporal constraints and uncertainty cannot be ignored remains very challenging. Currently, no single temporal planner exhibits strong performance and, at the same time, handles all the features needed to represent practical problems. This project aims to contribute to filling this gap. On the one hand, we will investigate how different representations of temporal planning problems impact the performances of existing planners and whether there is one representation that facilitates efficient and flexible reasoning. On the other hand, we will formulate efficient algorithms that support advanced features of temporal reasoning such as required concurrency, timed transitions and uncontrollable action durations.
The second strand of this project emerges from the observation that, in any complex real-world operations, artificial artefacts rarely operate in isolation from humans. For the humans and the agents to team up in a fluidly and trustworthy, it is crucial that the agents' decision-making is intelligible to the human operators and also receptive to inputs from them. In this project, we explore the idea that planning can play a pivotal role in achieving intelligibility in autonomous systems. We consider two different facets of intelligibility: ex-post intelligibility, or explainability, whereby the system can exhibit the information and the logic that it has used to arrive at its decisions; and ex-ante intelligibility, or transparency, whereby the system exposes how it operates to a human operator in such a way that the operator can intervene and negotiate with the system a different course of actions. We investigate how planning and computational argumentation can be blended to achieve both ex-post and ex-ante intelligibility. Argumentation refers to a set of techniques for evaluating claims by considering reasons for and against them through logical reasoning. Argumentation techniques based on planning will empower the agent with the capacity to exhibit arguments in support of its decisions as well as to negotiate with the operator a change in the plan if needed.
Providing advances in the planning and collaboration skills of autonomous systems would benefit research in planning, AI and robotics and, more crucially, promote their broad adoption in real-world contexts.
The primary objective of the project is broken down into two strands. The first is to equip artificial artefacts that operate in real-world settings with the ability to reason about themselves and the world around them to determine plans for achieving high-level goals efficiently and robustly. Planning is a key component of intelligence and one of the most traditional fields of artificial intelligence (AI). Planning has achieved impressive results in idealised settings where the world is deterministic, and actions are instantaneous. However, planning in real-world environments in which temporal constraints and uncertainty cannot be ignored remains very challenging. Currently, no single temporal planner exhibits strong performance and, at the same time, handles all the features needed to represent practical problems. This project aims to contribute to filling this gap. On the one hand, we will investigate how different representations of temporal planning problems impact the performances of existing planners and whether there is one representation that facilitates efficient and flexible reasoning. On the other hand, we will formulate efficient algorithms that support advanced features of temporal reasoning such as required concurrency, timed transitions and uncontrollable action durations.
The second strand of this project emerges from the observation that, in any complex real-world operations, artificial artefacts rarely operate in isolation from humans. For the humans and the agents to team up in a fluidly and trustworthy, it is crucial that the agents' decision-making is intelligible to the human operators and also receptive to inputs from them. In this project, we explore the idea that planning can play a pivotal role in achieving intelligibility in autonomous systems. We consider two different facets of intelligibility: ex-post intelligibility, or explainability, whereby the system can exhibit the information and the logic that it has used to arrive at its decisions; and ex-ante intelligibility, or transparency, whereby the system exposes how it operates to a human operator in such a way that the operator can intervene and negotiate with the system a different course of actions. We investigate how planning and computational argumentation can be blended to achieve both ex-post and ex-ante intelligibility. Argumentation refers to a set of techniques for evaluating claims by considering reasons for and against them through logical reasoning. Argumentation techniques based on planning will empower the agent with the capacity to exhibit arguments in support of its decisions as well as to negotiate with the operator a change in the plan if needed.
Providing advances in the planning and collaboration skills of autonomous systems would benefit research in planning, AI and robotics and, more crucially, promote their broad adoption in real-world contexts.
Planned Impact
The goal of this project is to equip artificial artefacts, such as robots and software agents, with the ability to autonomously undertake sophisticated tasks and, when needed, to collaborate with humans in a natural and mutual intelligible way.
The potential for society and economy of robotics and autonomous systems is well understood and has been recently highlighted by many studies. The report "Disruptive technologies" by McKinsey identifies in advanced robotics and autonomous systems two of the main technologies that "have the potential to truly reshape the world in which we live and work" and estimates that advanced robotics could generate a potential impact of $1.9 - 6.4 trillion per year by 2025. The Royal Society report "Robotics and autonomous systems" states: "Robotics and autonomous are of immense societal impact, pervading all areas of society including medicine, transport, and manufacturing."
The use of both software agents and robotic artefacts in everyday life has risen sharply, and we see increasingly more examples of their use in society with several commercial products already on the market. As interaction with humans increases so does the demand for sophisticated capabilities associated with deliberation and high-level cognitive functions. WP1 of my proposal responds to this need and deals with endowing artificial systems with higher level cognitive functions that enable them to reason and act in complex environments. Planning is, in fact, considered central in the spectrum of capabilities required for autonomy.
The Royal Society report also states: "Fully autonomous robotics can be problematic. As a result, there is a shift from isolated decision-making systems to those that share control, with significant autonomy devolved to robotics and autonomous systems, leaving end-users to make only high-level decisions. Shared autonomy will demand the closing of the semantic gap between human and machine." WP2 of my proposal deals with shared autonomy and aims at contributing to close "gap between human and machine" by combining planning and argumentation to give rise to fluid and transparent teaming between human operators and artificial artefacts. The use of these two techniques will empower the human part with the ability to enquire into the behaviour of the machine and directly argue with it to better understand its line of reasoning and change it, if appropriate.
The need for an intelligible AI has gained a lot of traction in academia, industry and the public sector recently, and has been highlighted by several official reports. The House of Lords report "AI in the UK: ready, willing and able?" states: "We believe that the development of intelligible AI systems is a fundamental necessity if AI is to become an integral and trusted tool in our society." The techniques at the core of this proposal will help in achieving more intelligible autonomous agents, impacting a variety of AI artefacts, from virtual characters that interact with users online to autonomous cars to service robots that assist people at home.
WP3 deals with an application of clear importance for society: supporting disaster response operations via autonomous UAVs. Humanitarian disasters cost lives and can cause huge setbacks. The UK Government's Humanitarian Policy emphases how the use of innovative technology can have a considerable impact on improving the efficiency and reach of humanitarian assistance. When a disaster strikes, UAVs can provide support with risk assessment, mapping, planning and search-and-rescue in the affected region. Currently, two of the main barriers to the use of UAVs for disaster relief are that the availability of expert pilots to teleoperate them and the lack of transparency in the behaviour of these UAVs. This project deals with both these problems as it aims to develop autonomous UAVs that do not require to be continuously piloted and whose conduct is decided in concert with the human operators.
The potential for society and economy of robotics and autonomous systems is well understood and has been recently highlighted by many studies. The report "Disruptive technologies" by McKinsey identifies in advanced robotics and autonomous systems two of the main technologies that "have the potential to truly reshape the world in which we live and work" and estimates that advanced robotics could generate a potential impact of $1.9 - 6.4 trillion per year by 2025. The Royal Society report "Robotics and autonomous systems" states: "Robotics and autonomous are of immense societal impact, pervading all areas of society including medicine, transport, and manufacturing."
The use of both software agents and robotic artefacts in everyday life has risen sharply, and we see increasingly more examples of their use in society with several commercial products already on the market. As interaction with humans increases so does the demand for sophisticated capabilities associated with deliberation and high-level cognitive functions. WP1 of my proposal responds to this need and deals with endowing artificial systems with higher level cognitive functions that enable them to reason and act in complex environments. Planning is, in fact, considered central in the spectrum of capabilities required for autonomy.
The Royal Society report also states: "Fully autonomous robotics can be problematic. As a result, there is a shift from isolated decision-making systems to those that share control, with significant autonomy devolved to robotics and autonomous systems, leaving end-users to make only high-level decisions. Shared autonomy will demand the closing of the semantic gap between human and machine." WP2 of my proposal deals with shared autonomy and aims at contributing to close "gap between human and machine" by combining planning and argumentation to give rise to fluid and transparent teaming between human operators and artificial artefacts. The use of these two techniques will empower the human part with the ability to enquire into the behaviour of the machine and directly argue with it to better understand its line of reasoning and change it, if appropriate.
The need for an intelligible AI has gained a lot of traction in academia, industry and the public sector recently, and has been highlighted by several official reports. The House of Lords report "AI in the UK: ready, willing and able?" states: "We believe that the development of intelligible AI systems is a fundamental necessity if AI is to become an integral and trusted tool in our society." The techniques at the core of this proposal will help in achieving more intelligible autonomous agents, impacting a variety of AI artefacts, from virtual characters that interact with users online to autonomous cars to service robots that assist people at home.
WP3 deals with an application of clear importance for society: supporting disaster response operations via autonomous UAVs. Humanitarian disasters cost lives and can cause huge setbacks. The UK Government's Humanitarian Policy emphases how the use of innovative technology can have a considerable impact on improving the efficiency and reach of humanitarian assistance. When a disaster strikes, UAVs can provide support with risk assessment, mapping, planning and search-and-rescue in the affected region. Currently, two of the main barriers to the use of UAVs for disaster relief are that the availability of expert pilots to teleoperate them and the lack of transparency in the behaviour of these UAVs. This project deals with both these problems as it aims to develop autonomous UAVs that do not require to be continuously piloted and whose conduct is decided in concert with the human operators.
Organisations
- Royal Holloway University of London (Lead Research Organisation)
- Polytechnic University of Turin (Collaboration)
- National Aeronautics and Space Administration (NASA) (Collaboration)
- HARVARD UNIVERSITY (Collaboration)
- University of Toronto (Collaboration)
- Massachusetts Institute of Technology (Collaboration)
Publications
Piacentini C
(2019)
Autonomous Target Search with Multiple Coordinated UAVs
in Journal of Artificial Intelligence Research
Bernardini S.
(2020)
Through the lens of sequence submodularity
in Proceedings International Conference on Automated Planning and Scheduling, ICAPS
Bernardini S
(2020)
Through the Lens of Sequence Submodularity
Bernardini S.
(2020)
An optimization approach to robust goal obfuscation
in 17th International Conference on Principles of Knowledge Representation and Reasoning, KR 2020
Bernardini S
(2020)
An Optimization Approach to Robust Goal Obfuscation
Keren S.
(2020)
Reasoning about plan robustness versus plan cost for partially informed agents
in 17th International Conference on Principles of Knowledge Representation and Reasoning, KR 2020
Bernardini SB
(2021)
A Unifying Look at Sequence Submodularity
in Artificial Intelligence (AIJ)
Bernardini S
(2021)
A unifying look at sequence submodularity
in Artificial Intelligence
Bernardini SB
(2022)
A Network Flow Interpretation of Robust Goal Legibility in Path Finding
Description | During this award, I have worked on robust planning for autonomous missions in critical environments. I have developed a suite of AI techniques to underpin the autonomous behaviour of teams of UAVs (Unammaned Aerial Vehicles) that are in charge of search-and-tracking applications. These drones are capable of coordinating their behaviour to search for an evasive target on a road network. The techniques that I developed can be used in a variety of surveillance and search and rescue missions. Starting from the formulation of search-and-tracking problems, I found out that several problems in other areas, for example, scheduling, recommender systems, online advertising, can be abstractly modelled in the same way. I have then devised a generalised algorithm that can solve all these problems in a more efficient way than related techniques. I have also formulated new techniques that combine AI planning and computational argumentation to allow a human operator to understand and change a plan if needed. |
Exploitation Route | Since the start of this award, I have been already successful in securing other four grants (three funded by Innovate Uk and one by the Leverhulme Trust) that build directly on the work carried out during the award. I have also submitted two addition grant proposals, one to EPSRC and one to the ORCA Hub, that focus on themes emerging from the SHARPA award. |
Sectors | Aerospace Defence and Marine Digital/Communication/Information Technologies (including Software) |
Description | The search-and-tracking techniques that I have developed during this award can be directly applied in surveillance and search and rescue missions as well as in multiple applications in extreme environments. Autonomous operations in such environments are expected to have an enormous societal and economic impact. Currently, I am using techniques emerged during the SHARPA award to tackle sophisticated missions in three domains, nuclear decommissioning, offshore energy production and mining, in the context of three Innovate UK projects. I have been working with stakeholders and industries to bring these techniques on the market as soon as possible. During the SHARPA award, I have organised a Summer School in Cognitive Robotics that was attended by more than 70 students. The school proved to be instrumental in encouraging students to become researchers in AI and robotics and focus on the challenges that remain open to bringing autonomous technology in hazardous environments, where it can produce a step-change in ensuring safe and effective operations. |
First Year Of Impact | 2019 |
Sector | Aerospace, Defence and Marine,Digital/Communication/Information Technologies (including Software),Security and Diplomacy |
Impact Types | Societal Economic Policy & public services |
Description | Connect-R |
Amount | £6,000,000 (GBP) |
Organisation | Innovate UK |
Sector | Public |
Country | United Kingdom |
Start | 01/2019 |
End | 02/2021 |
Description | Game theory for large-scale, hybrid decision-making problems |
Amount | £28,000 (GBP) |
Organisation | The Leverhulme Trust |
Sector | Charity/Non Profit |
Country | United Kingdom |
Start | 12/2019 |
End | 12/2020 |
Description | MIMRee |
Amount | £4,200,000 (GBP) |
Organisation | Innovate UK |
Sector | Public |
Country | United Kingdom |
Start | 03/2019 |
End | 02/2021 |
Description | Prometheus |
Amount | £2,000,000 (GBP) |
Organisation | Innovate UK |
Sector | Public |
Country | United Kingdom |
Start | 03/2019 |
End | 12/2021 |
Title | Generalised Greedy Approach |
Description | This is a new algorithm to generate sequence of items that maximise a sub-modular objective function. |
Type Of Material | Computer model/algorithm |
Year Produced | 2020 |
Provided To Others? | Yes |
Impact | This technique can be used in many different fields, from automated planning to scheduling to recommender systems. |
Title | Search and tracking |
Description | A set of techniques to underpin the autonomous behaviour of a team of UAVs involved in search and tracking applications. |
Type Of Material | Computer model/algorithm |
Year Produced | 2019 |
Provided To Others? | Yes |
Impact | This suite of techniques can be used in several surveillance missions and search and rescue missions. |
Description | Harvard university |
Organisation | Harvard University |
Department | School of Engineering and Applied Sciences |
Country | United States |
Sector | Academic/University |
PI Contribution | My partner and I worked together on reasoning about plan robustness and plan cost for partially informed autonomous agents. |
Collaborator Contribution | My partner and I worked together on reasoning about plan robustness and plan cost for partially informed autonomous agents. |
Impact | My partner and I have submitted a paper to the 17th International Conference on Principles of Knowledge Representation and Reasoning titled "Reasoning About Plan Robustness Versus Plan Cost for Partially Informed Agents". |
Start Year | 2019 |
Description | MIT collaboration |
Organisation | Massachusetts Institute of Technology |
Country | United States |
Sector | Academic/University |
PI Contribution | The contributions that I made to this collaboration are several and of different nature. I have provided the MIT team with intellectual input regarding how to perform a translation between different languages used to support autonomy in robots. I have participated in the training of MIT staff by working with PhD and Master students. I have also provided access to the code that I wrote in the past to facilitate the full implementation of the above-mentioned translation. |
Collaborator Contribution | My collaborators at MIT provided me with access to the MIT's facilities as well as their code to underpin autonomy in robotic artefacts. The head to the group with who I collaborate has also provided me with intellectual input and guidance to reach the goals of my research. |
Impact | 1) New/Improved Technique/Technology - Modelling languages for automated planning domains (2019) 2) Software - Translation from PDDL to Concurrent Automata (2019) |
Start Year | 2018 |
Description | NASA collaboration |
Organisation | National Aeronautics and Space Administration (NASA) |
Country | United States |
Sector | Public |
PI Contribution | I have provided NASA with my expertise and intellectual input concerning how to model manipulation tasks that a robotic arm needs to perform on the International Space Station as an automated planing problem. |
Collaborator Contribution | NASA has provided me with access to their data and models, which has enriched my understanding of the real-world problems that need to be tackled via automated planning. |
Impact | Planning models of the manipulation tasks that a robotic arm needs to perform on the International Space Station. |
Start Year | 2018 |
Description | Politecnico of Turin |
Organisation | Polytechnic University of Turin |
Country | Italy |
Sector | Academic/University |
PI Contribution | My partner and I worked together on some aspect to generating sequences of items (i.e. actions) efficiently. This is relevant in planning, scheduling and related fields. |
Collaborator Contribution | My partner and I worked together on some aspect to generating sequences of items (i.e. actions) efficiently. This is relevant in planning, scheduling and related fields. |
Impact | The following paper has been prepared and accepted for publication: Through the Lens of Sequence Submodularity Sara Bernardini, Fabio Fagnani and Chiara Piacentini Proceedings of the 30th International Conference on Automated Planning and Scheduling (ICAPS-20). Nancy, France, June 2020. This paper has received the ICAPS-2020 Best Paper Honorable Mention Award. |
Start Year | 2018 |
Description | University of Toronto |
Organisation | University of Toronto |
Country | Canada |
Sector | Academic/University |
PI Contribution | My partner and I worked together on generating sequences of items (e.g. actions) efficiently. This is relevant in planning, scheduling and related fields. We have also worked together on robust planning for search and tracking missions. |
Collaborator Contribution | My partner and I worked together on generating sequences of items (e.g. actions) efficiently. This is relevant in planning, scheduling and related fields. We have also worked together on robust planning for search and tracking missions. |
Impact | The following paper has been prepared and accepted for publication: Through the Lens of Sequence Submodularity Sara Bernardini, Fabio Fagnani and Chiara Piacentini Proceedings of the 30th International Conference on Automated Planning and Scheduling (ICAPS-20). Nancy, France, June 2020. The paper has received the ICAPS-2020 Best Paper Honorable Mention Award. We have also published the following paper in one of the most established journals in AI: Autonomous Target Search with Multiple Coordinated UAVs Chiara Piacentini , Sara Bernardini and Chris Beck Journal of Artificial Intelligence Research (JAIR). Volume 65, Pages 519-568, August 2019. |
Start Year | 2018 |
Title | Generalised Greedy Algorithm |
Description | This is an implementation of our new technique to find sequences of items via a generalised greedy approach. |
Type Of Technology | Software |
Year Produced | 2019 |
Impact | This algorithm can be used in many different fields, from planning to scheudling to recommender systems. |
Title | Modelling languages for automated planning domains |
Description | There are several ways to describe planning domains. Over the years, different communities have developed their own techniques, each one capable of capturing a subset of the typical characteristics of planning domains. My work investigates ways to translate one modelling language into another and explores under which circumstances it is best to use each particular language. |
Type Of Technology | New/Improved Technique/Technology |
Year Produced | 2019 |
Impact | Correct translation between different planning domain specification languages impact the entire planning community as they facilitate the cross-fertilisation of ideas between different groups and allow one sub-community to use and test the tools developed by another one. My translation is used at MIT and NASA, for example. |
Title | Multi-UAV search and tracking |
Description | This is the implementation of our suite of AI techniques for underpinning the behaviour of multiple UAVs tasked with search-and-tracking missions. |
Type Of Technology | Software |
Year Produced | 2019 |
Impact | This software could be used to operate multiple UAVs tasked with search-and-tracking missions. |
Title | Translation from PDDL to Concurrent Automata |
Description | This software implements a translation from PDDL (Planning Domain Definition Language) into Concurrent Automata. These are two different methods to encode planning domains. |
Type Of Technology | Software |
Year Produced | 2019 |
Impact | The main impact of this software is that it allows researchers to use different planning technologies to solve the same planning problem and obtain different solutions. |
Description | AAAI-23: Thirty-Seventh AAAI Conference on Artificial Intelligence. |
Form Of Engagement Activity | Participation in an activity, workshop or similar |
Part Of Official Scheme? | No |
Geographic Reach | International |
Primary Audience | Other audiences |
Results and Impact | I was the Associate General Chair of the Conference |
Year(s) Of Engagement Activity | 2023 |
URL | https://aaai-23.aaai.org/ |
Description | BBC Interview |
Form Of Engagement Activity | A press release, press conference or response to a media enquiry/interview |
Part Of Official Scheme? | No |
Geographic Reach | National |
Primary Audience | Media (as a channel to the public) |
Results and Impact | This was an interview at the BBC in which the interviewer asked me about autonomous systems in offshore energy. |
Year(s) Of Engagement Activity | 2019 |
Description | Debate: Fear and AI: Will I be replaced? |
Form Of Engagement Activity | A talk or presentation |
Part Of Official Scheme? | No |
Geographic Reach | National |
Primary Audience | Public/other audiences |
Results and Impact | This activity was a debate on the theme Fear and AI: Will I be replaced? It was organised by StoryFuture Academy. |
Year(s) Of Engagement Activity | 2019 |
Description | Guest Editor of the Journal "Artificial Intelligence" for the Special Issue "Risk-Aware Autonomous Systems: Theory and Practice" |
Form Of Engagement Activity | Participation in an activity, workshop or similar |
Part Of Official Scheme? | No |
Geographic Reach | International |
Primary Audience | Other audiences |
Results and Impact | This is a special issue for one of the most important journals in AI, the journal of Artificial Intelligence (AIJ), on the topic "Risk-Aware Autonomous Systems: Theory and Practice". Academics and professionals who work on this topic will have a chance to contribute to this special issue, which will be published at the end of 2021. |
Year(s) Of Engagement Activity | 2021 |
Description | Participation in ICAPS |
Form Of Engagement Activity | Participation in an activity, workshop or similar |
Part Of Official Scheme? | No |
Geographic Reach | International |
Primary Audience | Other audiences |
Results and Impact | I presented my research about autonomous agents. I was the Workshop Chair so I made sure that the workshops were running smoothly. |
Year(s) Of Engagement Activity | 2019 |
Description | Summer School in Cognitive Robotics |
Form Of Engagement Activity | Participation in an activity, workshop or similar |
Part Of Official Scheme? | No |
Geographic Reach | International |
Primary Audience | Postgraduate students |
Results and Impact | I am organising the a Summer School in Cognitive Robotics for undergraduate/postgraduate students and for interested professionals. The school aims at forming the next generation of researchers/practitioners in intelligent, autonomous robots. |
Year(s) Of Engagement Activity | 2019 |