Challenge 3: Sustainability through automation

Climate change is the grand challenge of our age. Automation has an important role to play here by improving the efficiency of polluting industries (thus reducing resource consumption and greenhouse gas emissions) as well as through enabling next-generation energy and mobility systems.
Our sustainability work is organised under four themes. In the first, mobility, we are developing better control techniques to tackle the vast challenges inherent in designing the resource-efficient and accessible mobility that is urgently needed for a sustainable future. These problems arise from the complexity of transport networks, but also from issues of accessibility, fairness and more.
The second theme tackles advanced manufacturing methods, which employ robotics, additive printing, digital twins and more. These technologies create opportunities for major improvements through better optimisation and control (in particular when it comes to adaptability, robustness and reliability); we are developing methodologies for process improvements, and for expanding beyond process to systems-level control that could bring great efficiencies in areas such as supply chain management.
In the third theme, energy systems, we address the challenges arising from the ongoing energy transition. As power generation becomes ever more decentralised and complex, with consumers becoming prosumers through photovoltaic installations, and incorporates a growing proportion of renewable energy sources, there is an urgent need for automation solutions to coordinate supply and demand, ensuring reliable supply despite the increased uncertainty and complexity in the system.
Finally, we are pursuing advanced control for robotics, aiming to develop methods that can enable autonomous agents to behave in dynamic environments, process contextual commands, interact with other agents, and so on. To do this, we have set ourselves the challenge of combining model predictive control with deep neural networks.
Theme 3.1: Mobility
Theme lead: Nikolas Geroliminis
Resource-efficient and accessible mobility is crucial to a sustainable future – not just because of the impact of emissions, but because of its key role in other sustainability aspects, such as access to transportation. Emerging technologies and transport modes (from ride-hailing services to autonomous vehicles) add complexity to the urgent task of managing complex transport infrastructure in rapidly expanding cities. When you factor in issues of adaptability, accessibility and fairness (including problems around data sharing and privacy, as mentioned above), the scale and delicacy of the challenge is clear.
We need better theoretical understanding to tackle these problems. We are building on and improving various control techniques, from hierarchical/distributed control to game theory, to promote advanced dynamic strategies for complex mobility systems. We investigate how the operations of those dynamic systems can be improved through strategic decisions, to integrate uncertainties in demand, user behaviour, congestion and energy availability.
Thread A: New paradigms in inter-modal and multi-usage mobility
Contributors: A. Censi, F. Corman, E. Frazzoli, N. Geroliminis, J. Lygeros
We are developing mathematical models, methods and simulations for ride-sharing systems that allow riders to switch between modes or drivers at a designated transfer hub – addressing the problem of the need for similar itineraries between matched drivers and riders. We are also looking at developing policies to manage rapidly changing demands in transport networks (i.e., bridging the gap between reality and the current state of the theory, which is based on static or equilibrium conditions).
Thread B: How to incorporate automation in the infrastructure planning for future mobility
Contributors: A. Censi, F. Corman, E. Frazzoli, N. Geroliminis, G. Hug
The communications capabilities baked into future mobility, combined with new control solutions, will open up new possibilities for infrastructure design. Meanwhile, the accelerating electrification of transport puts new pressure on the power grid. We tackle the problem of how to design vehicle charging infrastructure, taking into account a wide range of operational and planning factors.
Thread C: Crowd-shipping and logistics in urban areas
Contributors: A. Censi, E. Frazzoli, N. Geroliminis, D. Kuhn
Traditional last-mile delivery using large vans comes with congestion, parking and access problems. The proposed alternative of crowd-shipping relies on individuals (typically cyclists) making slight adjustments to their itineraries to collect and deliver packages. We are investigating transfer points (to reduce the match needed between short planned bike trips and long package journeys across a city), dynamic pricing strategies, and integration with wider transport systems (including ride-sharing as well as mass transit), and working toward a pilot study to test real-world applicability.
Theme 3.2: Advanced Manufacturing
Theme lead: Alisa Rupenyan
Today’s advanced manufacturing processes, which employ technologies such as additive manufacturing (3D printing) and robotics, offer rich opportunities for improved control systems. Models are needed at every level of abstraction, from partial differential equations at the process level, through discrete event models at the plant level, right through to flow models at the system level. Control in this field must provide robustness and adaptation to environmental changes, requirements, and disturbances, with data-driven approaches central to improving control methodologies, as well as better modelling systems for improved reliability.
We have already achieved significant advances in this field, especially at the process level – developing new methods for modelling, optimisation and feedback control, and demonstrating significant improvements in a wide range of processes as a result; digital twins (keeping them up to date, and using them as data generators) have formed an important part of this research. Beyond the process level, we have demonstrated optimisation and control approaches for scheduling and dispatch at different levels (system, process, networked systems) in a shop floor. Looking forward, we are addressing key research challenges that emerge from new game-changing concepts (such as individualised production, collaborative manufacturing and zero-defect manufacturing) that have yet to be formalised from a systems and control perspective.
Thread A: Process-level automation
Contributors: E. Balta, A. Karimi, J. Lygeros, A. Rupenyan
The complex physical and/or chemical phenomena inherent in many manufacturing processes pose an insurmountable challenge for pure model-based approaches, which may be too complex or inaccurate to be useful. We are using data-driven methods to fill the gap.
Thread B: From the process level to the system level
Contributors: E. Balta, J. Lygeros, A. Rupenyan
Beyond process-level control, we aim to develop design principles to, on one hand, abstract information from the process level to inform system-level decisions (demand forecasting, supply-chain management and so on), and on the other, to pass high-level control objectives to the process level. This involves game theory challenges such as competitive behaviour or information sharing. Our long-term goal is to achieve a holistic optimisation approach across levels within the manufacturing ecosystem, supporting improvements in sustainability along with other optimisation goals such as manufacturing cost and resilience.
Thread C: Digital twin technologies
Contributors: E. Balta, P. Heer, A. Rupenyan
Digital twins (representations of a system that change along with the real system, through a stream of data updates and adaptations) hold great potential for operational aspects such as predictive maintenance, planning and more. As the digital twin is adjusted using process data, the underlying control algorithm needs to be adapted. This provides an ideal testing ground for the research into lifelong learning and control described under Theme 1.4.
Theme 3.3: Energy Systems
Theme lead: Philipp Heer
In pursuing net-zero emissions, the energy sector is transitioning (at an ever-increasing rate) from a system oriented around central supply, and mostly based on fossil energy carriers, to a more decentralised and renewable system that comes with great increases in digitalisation, as well as associated challenges.
The new systems hold great potential for flexibility, advanced control, and previously untapped synergies of planning and operations. They also hold a much higher degree of variability than classical power generation (with photovoltaics in particular bringing weather dependence into play). That means that flexibility, and local balancing of energy provision from distributed resources, is of great importance; which demands general and adaptable schemes to coordinate supply and demand between consumers and distributed resources.
There is also a need to develop long-term strategies to manage the interdependence between energy system planning and operation, including the use of digital twins.
Thread A: Energy System Levels of Abstraction
Contributors: P. Heer, S. Mastellone, A. Rupenyan
The large-scale introduction of power electronics converters in energy and industrial processes has transformed these applications into highly dynamic, interactive systems. Since these technologies can serve diverse uses, with different temporal and spatial scales, it is useful to develop adaptive frameworks (with various levels of abstraction) that can be employed in various use cases. We aim to define a unified framework to address the complexity of regulating the power flow across the mechanical and electrical parts, to ensure the overall security of supply. On larger scales, we are looking to achieve better understanding of the exchange of relevant information from and to the lower levels, in order to abstract from relevant data with negligible information loss. This involves modular thinking that unifies multiple disciplines, from engineering to economics and policy.
Thread B: Coordination of Distributed Energy Resources
Contributors: P. Heer, G. Hug, C. Jones, J. Lygeros, S. Mastellone, V. Medici, D. Shaw
Switzerland already has the regulatory framework to support local coordination of energy prosumers, though this is not yet widely implemented. Through a so-called ZEV (Zusammenschluss zum Eigenverbrauch), neighbours can act (and be billed by the utility) as a single entity. Within this framework, we aim to (1) derive an approach to form building clusters that can minimise central balancing needs; (2) automate the coordination between houses, for efficiency and privacy; (3) define a mechanism to incentivise effective coordination between the clusters. We will also consider the arising ethical questions, such as the risk that incentives may favour certain already privileged groups, and whether markets are the fairest and most ethical mechanism for coordinating energy distribution.
Thread C: Co-synthesis of energy systems design and control over time
Contributors: A. Censi, E. Frazzoli, P. Heer, G. Hug
In planning a building’s energy systems, decisions must unavoidably be based on many assumptions and simplifications. At this early stage, of course no data has yet been collected on actual use; nor is it possible to foresee how needs may evolve over the 50-plus years of a building’s lifespan. We aim to develop efficient computational tools to improve decision making (taking into account flexibility and redundancy measures to ensure reliability in a highly automated system), and crucially, to provide the transparency (now lacking in data-driven control techniques) that will enable operators to comprehend the reasoning behind automated processes, improving trust as well as troubleshooting.
Theme 3.4: Advanced Control for Robotics
Theme leads: Marco Hutter and Fisher Yu
Future automated systems, such as self-driving cars and home service robots, need to learn to behave in dynamic environments, process contextual commands, interact with other agents, etc – constraints and procedures that are hard, or even impossible, to formulate precisely. We are pursuing the ideal of a general algorithm for such scenarios, drawing on the success of learning algorithms that can capture complex relationships from large amounts of experience, instead of relying on manual designs.
We aim to mimic human locomotion behaviour and to learn mappings from input signals to output motion in an end-to-end fashion, such that the whole system can learn and adapt from offline and online experience without the involvement of human engineers. Our main approach is to combine model predictive control (a strong area within the NCCR Automation) with deep representative learning, a well-studied approach in machine learning, computer vision, and natural language processing. This however involves certain challenges. While MPC assumes it is possible to find (near) optimal solutions quickly, the models represented by deep neural networks are highly non-linear. Further, traditional models have a negligible number of parameters compared with deep models.
Thread A: Deep Model Predictive Control for Robotics
Contributors: M. Hutter, C. Jones, M. Zeilinger
Legged locomotion control over rough terrain presents major challenges, owing to the complexity and uncertainty of the system dynamics as well as limited computational power. Combining the complementary approaches of MPC and reinforcement learning, we aim to develop perception-aware controllers (with a vision-based MPC helping to compensate for sensor limitations), test them using the quadrupedal robot ANYmal available at ETH Zurich, and benchmark the work in simulation and real-world experiments against existing approaches.
Thread B: Human-Robot Interaction
Contributors: A. Censi, E. Frazzoli, P. Heer, D. Shaw, M. Zeilinger
As machine learning proliferates, we continue to wrestle with the dilemma of how to evaluate machine-made decisions, which lack transparency and explainability. For instance, we urgently need to be able to understand decisions made by autonomous vehicles, for legal and liability reasons. Building on previous work at the NCCR Automation, we will investigate the contrasting approaches of a “rulebook” orientation (with decision making constrained by clearly defined hierarchical rules) or data-driven learning approach (which may better accommodate future uncertainties), with the goal of informing future developments in autonomous decision-making systems. We are also researching the ethics of interaction between humans and humanoid robots, specifically in the context of elderly care.
Thread C: Autonomous Robots
Contributors: A. Censi, E. Frazzoli, M. Hutter
The third robotic revolution will see the integration of mobile robotic systems in unstructured everyday environments, ultimately replacing humans in a range of dangerous and tedious jobs. We are investigating legged mobile manipulator systems for inspection and maintenance tasks: starting with end-to-end systems for navigation and locomotion, we will expand the agent’s skillset to combined locomotion and manipulation, using reinforcement learning to achieve unprecedented versatility and agility. The ultimate goal is to develop a multi-agent solution for coordinating a fleet of robots.
See related publications and data sets and news items.