Multiple optimization objectives arise frequently in robot motion planning. If improving one objective degrades another, the set of Pareto-optimal solutions (such that no one is clearly better than another) can be infinite. While scalarization or approximation methods are commonly used to address this challenge, exact computation of this set of solutions is possible for certain problems. Path planning that simultaneously optimizes length and clearance is one of such problems. In this talk, I will present a complete algorithm that computes the Pareto-optimal solutions that optimize path length and clearance and a data structure that finitely represents the set of all Pareto-optimal paths in plane. Particular optimal paths can then be selected from the computed data structure during execution, based on any additional conditions or considerations.

Robots are deployed in an increasing number of environments and applications, and tasked with ever complex missions with temporal, logical, and timing constraints. Practical and safety considerations impose that solutions are robust to perturbations and noise, and able to handle infeasible scenarios as best as possible. A planner for a self-driving car may not just return without a solution, it still needs to steer the car as best possible. A multi-agent team should not abandon a mission if one robot fails during deployment. A delivery robot should find alternatives and swap unavailable groceries based on the user's preferences. In this talk, we present and clarify the robustness and relaxation of temporal logic specifications when dealing with multiple objectives. We apply these techniques to coordinate heterogeneous teams of robots, robot swarms, and modular aerial robotic systems.

Robotic tasks involve multiple objectives that need to be jointly optimized. In many tasks, humans are good at making the tradeoff between those objectives. So how can we learn the tradeoffs from humans' supervision? First, I will discuss preference-based reward learning as a popular approach for this problem. I will then discuss the issues around this approach, mainly time- and data-efficiency, as well expressivity. As a solution, I am going to propose learning the reward functions that encode the tradeoffs using language feedback from the humans. Specifically, I will present one approach for learning reward functions from "comparative" language feedback which addresses the drawbacks of preference-based learning. 

When designing complex systems, we need to consider multiple trade-offs at various abstraction levels and scales, and choices of single components need to be studied jointly. For instance, the design of future mobility solutions (e.g., autonomous vehicles, micromobility) and the design of the mobility systems they enable are closely coupled. Indeed, knowledge about the intended service of novel mobility solutions would impact their design and deployment process, while insights about their technological development could significantly affect transportation management policies. Optimally co-designing sociotechnical systems is a complex task for at least two reasons. On one hand, the co-design of interconnected systems (e.g., large networks of cyber-physical systems) involves the simultaneous choice of components arising from heterogeneous natures (e.g., hardware vs. software parts) and fields, while satisfying systemic constraints and accounting for multiple objectives. On the other hand, components are connected via collaborative and conflicting interactions between different stakeholders (e.g., within an intermodal mobility system). In this talk, I will present a framework to co-design complex systems, leveraging a monotone theory of co-design and tools from game theory. The framework will be instantiated in the task of designing future mobility systems, all the way from the policies that a city can design, to the autonomy of vehicles as part of an autonomous mobility-on-demand service. Through various case studies, I will show how the proposed approaches allow one to efficiently answer heterogeneous questions, unifying different modeling techniques and promoting interdisciplinarity, modularity, and compositionality. I will then discuss open challenges for compositional systems design optimization, and present my agenda to tackle them.

This talk focuses on preference-aware planning for robotic systems and its connection to multi-objective planning. It reviews foundational concepts in multi-objective planning, including various notions of optimality such as Pareto optimality and scalarization-based optimality. The talk illustrates how different families of preference-based planning problems can be viewed as special cases of multi-objective planning, where users assign preferences over multiple, often conflicting, objectives, indicating their relative importance. Two general settings are considered: One in which all the objectives are comparable, leading to a total order, and another in which some objectives are incomparable, inducing a preorder. The talk uses three notions from the stochastic order theory—namely, strong, weak, and weak* stochastic orderings—to show how different policies (plans) are ranked in stochastic environments. Using these orderings, the preference-based planning problem in stochastic settings is reduced to computing Pareto-optimal policies for a transformed multi-objective Markov Decision Process (MDP), where the objectives are derived through set-based constructions from the original preference structure and the objectives. Examples from recent work on planning with preferences over temporally extended goals, in both deterministic and stochastic environments, are presented to demonstrate the applicability of these concepts and methods.

Autonomous systems often operate in environments where multiple, potentially conflicting objectives must be balanced—such as safety, efficiency, and human comfort. In this talk, I will present a framework for multi-objective planning with contextual preferences, where the relative importance of objectives changes across different regions of the state space. I will discuss the applications of it in single and multi-agent settings to avoid negative side effects.

Bi-objective (and multi-objective) search algorithms allow the cost of every graph edge to be quantified by two (or more) real values. They essentially assume that one wants to find the set of all paths, called the Pareto frontier, such that each path in the set is better than all other paths from a given start vertex to a given goal vertex with respect to the sum of at least one cost component of its edges. In this talk, I will describe recent algorithmic progress on multi-objective path planning by a team of researchers from artificial intelligence and robotics that develops optimal and approximately-optimal bi-objective search algorithms by finding synergies between ideas from existing bi- and multi-objective search algorithms and recent algorithmic developments in the artificial intelligence search community.