Algorithms and Architectures for Improving Optimization Layers

anonymous@undisclosed.example.com (Anonymous) — Tue, 03 Dec 2024 23:02:29 +0000

Algorithms and Architectures for Improving Optimization Layers

Abstract:

Many real-world challenges, from robotic control to resource management, can be effectively formulated as optimization problems. Recent advancements have focused on incorporating these optimization problems as layers within deep learning pipelines, enabling the explicit inclusion of auxiliary constraints or cost functions, which is crucial for applications such as enforcing physical laws, ensuring safety constraints, and optimizing complex objectives. However, these layers introduce several challenges, including inference and representational inefficiencies, unstable/slow training dynamics, and modeling inaccuracies, which need to be addressed to fully harness their potential.

We systematically investigate these challenges and propose novel numerical methods and architectural solutions that mitigate them, making optimization layers more efficient and effective within deep learning pipelines. Our contributions include methods for enhancing computational efficiency, tackling issues of gradient bias and variance, and improving sample efficiency in reinforcement learning using approximate simulators. We demonstrate these contributions across different applications, ranging from input-optimization problems to SLAM/VO to reinforcement learning. We also present a new approach for visual-inertial navigation in nanosatellites, highlighting the practical benefits of integrating optimization layers in challenging real-world scenarios. Finally, we propose future work on extending our analysis and methods for improved representational fidelity and training stability to constrained optimization layers, using differentiable model predictive control layers as an example use case to illustrate their effectiveness.

Together, these contributions advance our understanding of the complexities and opportunities in integrating optimization layers within deep learning models, offering new frameworks and insights that improve efficiency, stability, and generalizability across a wide range of complex tasks.

SLAM, optimization, robotics, CMU, Swaminathan-Gurumurthy, 2024

Stanford Seminar - The Trouble with Contact: Helping Robots Touch the World

anonymous@undisclosed.example.com (Anonymous) — Sat, 07 Dec 2024 00:26:21 +0000

Stanford Seminar - The Trouble with Contact: Helping Robots Touch the World

Abstract:

Contact with the outside world is challenging for robots due to its inherently discontinuous nature – when a foot or hand is touching a surface the forces are completely different than if it is just above the surface. However, most of our computational and analytic tools for planning, learning, and control assume continuous (if not smooth or even linear) systems. Simple models of contact make assumptions (like plasticity and coulomb friction) that are known to not only be wrong physically but also inconsistent. In this talk I will present techniques for overcoming these challenges in order to adapt smooth methods to systems that have changing contact conditions. In particular I will focus on three topics: First, I will present the “Salted Kalman Filter” for state estimation over hybrid systems. Second, I will present an analysis approach that unifies and extends different strategies for stabilizing and controlling systems through contact. Finally, I will talk about when these hybrid models of contact break down, especially when driving on sand.

Playlist

Book Advanced Robot Dynamics - Impacts - Aaron M. Johnson

Book Motivation

“I developed and teach a graduate class called Robot Dynamics and Analysis. Unlike most other robot dynamics classes, this new class focused on contact, impacts, and how to handle changing contact conditions in simulation and optimization. However, while the first half used a standard textbook, I could not find any good educational resources that covered hybrid systems, changing contact conditions, and impacts at the level needed for this class. Instead, I used two of my own research papers and a collection of other sources. The problem with these resources are that: 1) research papers are not written like textbooks, and generally lack the detailed explanations and example problems necessary; and 2) they lack a coherent dialog and notation, as each was written by a different author with a different audience in mind.

To remedy these issues, I am writing a textbook to cover the topics listed below over the course of several years. This will allow for a cohesive treatment of contact and related topics in a form that will be more approachable for students, following research-based best practices for textbook design. The text will be based on my papers, class notes, and translation guides, which comments back from students indicated were more helpful than many of the assigned readings. The textbook will be available for free online, and I will solicit input from the community during its development to help ensure it is of use to a wider range of researchers”

Stanford, robotics, contact-dynamics, CMU, Aaron-Johnson, 2024

𝐶𝑎𝑟𝑎𝑚𝑖𝑧𝑎𝑟𝑢'𝑠 𝐵𝑙𝑜𝑔 feed:2024:12:03

Algorithms and Architectures for Improving Optimization Layers

Algorithms and Architectures for Improving Optimization Layers

Stanford Seminar - The Trouble with Contact: Helping Robots Touch the World

Stanford Seminar - The Trouble with Contact: Helping Robots Touch the World