About
Frontiers
Frontiers
Geometry and Global Optimality
Geometry and Global Optimality
Many optimization problems in robotics are characterized by a non-convex landscape stemming from both the costs and constraints which makes the problems hard to solve optimally. Thankfully, specific problem structure often allows the use of tools from globally optimal and Riemannian optimization.
Many optimization problems in robotics are characterized by a non-convex landscape stemming from both the costs and constraints which makes the problems hard to solve optimally. Thankfully, specific problem structure often allows the use of tools from globally optimal and Riemannian optimization.
Speed and Scale
Speed and Scale
Real-time requirements and long-horizon planning impose stringent demands on the efficiency and scalability of deployed solvers. To meet these demands, solvers may exploit structure-specific sparsity, reuse prior computations through warm-starting, or exploit parallelism and in particular GPU acceleration.
Real-time requirements and long-horizon planning impose stringent demands on the efficiency and scalability of deployed solvers. To meet these demands, solvers may exploit structure-specific sparsity, reuse prior computations through warm-starting, or exploit parallelism and in particular GPU acceleration.
Differentiable Optimization
Differentiable Optimization
Model-based optimization and deep-learned architectures are both essential components of modern robotics pipelines. Differentiable optimization and simulation, respectively, enables their seamless integration, as optimization layers and simulators can be included in end-to-end learned pipelines and their parameters can be trained jointly with the neural network parameters to optimize a higher-level task.
Model-based optimization and deep-learned architectures are both essential components of modern robotics pipelines. Differentiable optimization and simulation, respectively, enables their seamless integration, as optimization layers and simulators can be included in end-to-end learned pipelines and their parameters can be trained jointly with the neural network parameters to optimize a higher-level task.
Optimization Through Contact
Optimization Through Contact
Capturing friction, collisions, impacts, and hysteresis contributes to more accurate predictions, but often at the cost of dealing with nonsmooth dynamical systems. Different approaches can handle such loss of regularity, exploiting problem structures, principled regularizations, or task-specific knowledge and learning.
Capturing friction, collisions, impacts, and hysteresis contributes to more accurate predictions, but often at the cost of dealing with nonsmooth dynamical systems. Different approaches can handle such loss of regularity, exploiting problem structures, principled regularizations, or task-specific knowledge and learning.
Target speakers and audience
Target speakers and audience
Our target speakers and panelists are drawn from communities that include:
Our target speakers and panelists are drawn from communities that include:
- Practitioners working on (or with) novel optimization algorithms and tools targeted towards (or relevant for) robotics.
- Researchers applying novel concepts from optimization theory to robotics problems (e.g. global optimality, end-to-end learning, non-smooth optimization).
Our target audience includes:
Our target audience includes:
- Researchers that want to expand their repertoire in terms of understanding / extending or analyzing existing optimization tools and algorithms.
- Researchers that develop optimization for robotics and wish to learn from, and bring their knowledge to, other fields to foster progress in robotics as a whole.