SEA Nets: Safe and Explainable Autonomous Networks of Distributed Learning Agents
Abstract
In this project we explore learning among a network of robotic agents who must perceiveand act in a physical world dominated by dynamic, uncertain phenomena. We propose afundamental re-think of the deep learning framework for multi-robot systems. We arguethat thecurrent big-data, big-model, deep network paradigm leads to monolithic modelsthat learn slowly, have poor composability, scant safety guarantees, and often make decisionsfor reasons that are uninterpretable to humans. These properties are ill-suited to therequirements of a network of robotic agents, which have to make fast, trustworthy inferencesbased on limited experience, and must give a human-interpretable trace of their decisions.We focus on three thrusts, each of which aims to overcome a different challenge in the currentdeep learning paradigm applied to a multi-robot context: Thrust I: Compact ComposableModels for Multi-Robot Learning, Thrust II: Strong Physical Priors for Explainability, andThrust III: Self-Assessing Safety with Learned Models. In Thrust I, we explore Liquid TimeConstant (LTC) networks, a new neural network architecture with greater similarity to biologicalneural networks than existing models, which holds strong promise for data-efficientlearning and out-of-distribution generalization, especially on dynamic control-oriented tasks.We focus on developing methods for a group of robotic agents to cooperatively train and actin a dynamic world using LTC networks. In Thrust II we explore learning architectures thatmodel physics, either explicitly through differentiable physics simulators, or implicitly byimposing physical laws in the learning architecture itself (such as in Neural Radiance Fields(NeRFs)). Our hypothesis is that strong physical priors will improve data-efficiency in learning,and will lead to human interpretability and explainability since the network#s decisionswill be based on a physically interpretable substrate. Again, we focus on the rarely consideredscenario of roboticagents using these strong physical priors for online collaboration inthe physical world. Lastly, in Thrust III, we consider safety guarantees for learned multiagentcontrol and planning algorithms. We specifically consider deep network reachabilitycomputations, which explicitly compute reachable sets for deep learned control systems toverify safety properties. Such deep net reachability algorithms have not yet been consideredin the context of learned multi-agent control and coordination. We will also develop methodsusing conformal prediction as a statistical tool to give probabilistic safety guarantees formulti-robot deep learned policies. These three thrusts are all united by a common scenarioof multiple robotic agents communicating over a wireless network to accomplish long-termcollaborative goals informed by a history of observed data from on-board sensors. We willuse our existing multi-rotor and autonomous vehicle beds to experimentally validate ourtheoretical and algorithmic results on robot hardware in real-world operational scenarios,including collaborative visual target search, and multi-robot persistent surveillance tasks.(Abstract approved for public release)
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 12, 2023
- Source ID
- N000142312354
Entities
People
- Mac Schwager
Organizations
- Office of Naval Research
- Stanford University
- United States Navy