Human-Machine Collaboration Programme
Human-Machine Collaboration Programme, supported by Amazon Web Services
The ORI has received funding from Amazon Web Services (AWS) to support three projects within the Oxford Research Pillar of the Human-Machine Collaboration Programme aimed to revolutionse research within the fields of artificial intelligence (AI), robotics and human-centred computing.
The Oxford Research Pillar tackles the following challenges
- Long term autonomy for service robots
- Large scale mixed initiative autonomy for logistics
- Human-robot shared autonomy
- Autonomy in blue light emergency services
- Human-plus-robot workplace
- Responsible robots for the digital economy
The following summarises ORI’s contributions and research aims for 3 of the research challenges
Project 1: Long-Term Autonomy for Service Robots
This project addresses the need for autonomous mobile robots to operate in service environments for, or alongside, humans. We are interested in providing enabling technologies which are able to support a wide range of robotic service applications, from customer service and stock management, through industrial inspection, to social care.
The project will study the AI and machine learning capabilities required to enable robust service robot operation (including manipulation, locomotion, navigation, interaction) over very long durations (months or years). The key to such performance is using the robot (or multiple robots) to gather data online, to learn from this in-situ experience, and exploit the results of learning to improve performance. We envisage this happening at different time scales, from rapid learning of new object models and physical interaction skills via demonstrations and self-supervision, to performance statistics gathered from weeks or months of long-term operation.
Representative publications from 2020/2021
- Rigter, M., Lacerda, B., & Hawes, N. (2021). Minimax Regret Optimisation for Robust Planning in Uncertain Markov Decision Processes. The Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI-21).
- Rigter, M., Lacerda, B., & Hawes, N. (2021). Risk-Averse Bayes-Adaptive Reinforcement Learning. arXiv:2102.05762.
Project 2: Large-Scale Mixed-Initiative Autonomy for Logistics
The project addresses the need for near-future large-scale logistics systems, as required in warehouses and ports. One particular challenge in such environments is that both humans and robots need to complete different parts of the roles, but for high efficiency, their actions need to be planned jointly. This requires both learnt models of human task capabilities (e.g. picking items in warehouses, operating loading cranes) and a planning framework capable of assigning/scheduling tasks across potentially hundreds of humans and robots. These algorithms should be robust against the uncertainty present in modelling and instructing humans, and able to provide guarantees of quality-of-service for the entire system. A second challenge is the need to be able to quickly provide robots in logistics settings with new physical interaction skills when new objects or processes are introduced into the setting. Such a process should require minimal human intervention in order to not impact the overall throughput of the system.
Project 3: Human-Robot Shared Autonomy
This project addresses the need for robots which operate in hazardous, or poorly modelled task environments, where it is necessary or desirable for a human operator to take control of the robot platform to complete parts of the task. Examples include sharing control of an autonomous vehicle, demolition of a structure, inspection of a collapsed building, or drilling for oil. The key problem is for the AI system to plan to take advantage of the performance of the human, and predict/manage human inputs. The human could be called upon to perform parts of the task that are either high risk (e.g. driving a vehicle around some road works) or hard for the robot to complete (e.g. due to limited sensing). The robot could then complete the more mundane or predictable parts of the task.
Representative publications from 2020/2021
- Hung, C.M., Sun, L., Wu, Y., Havoutis, I., & Posner, I. (2021) Introspective Visuomotor Control: Exploiting Uncertainty in Deep Visuomotor Control for Failure Recovery. International Conference on Robotics and Automation (ICRA)
AIBotics Go-Digital Series - Nick Hawes - Long-Term and Large-Scale autonomy