Proactive Decision Making for Autonomous Systems: a Formal Methods Approach
Abstract
This proposal addresses the problem of proactive decision making for autonomous systems (AS) operating in shared environments, whereby an AS (1) proactively interacts with other agents (either human or robotic) to infer their intents while concurrently (2) exploiting this information to take actions that account for possible agent responses. Currently, AS tend to passively interact with their environment: they unobtrusively observe and make predictions about the future behavior of other agents, and plan as if the actions of other agents are unaffected by the AS s own actions. This oftentimes leads to defensive and potentially unsafe behaviors, corresponding to AS being ~surprised~ by theirsurroundings, as recent accidents involving self-driving cars have shown. The key insight behind this proposal is that an AS operating in a shared environment should proactively interact with other agents in order to elicit informative responses. For example, a self-driving car could carefully nudge into an adjacent lane to gauge the reactions of other human-driven vehicles and exploit such information to plan a lane change. Enabling proactive decision making requires significant strides in two key areas, namely (1) modeling, that is,formalisms to describe potentially complexinteractions among decision-making agents, and (2) decision theory, that is, decision-making strategies withinformation- and game-theoretic aspects todisambiguate intents (that other agents may try to purposefully obfuscate) and which can anticipate best responses in a probabilistic context. Leveraging recent advances in probabilistic formal methods, my technical approach will be to model the interaction between an AS and other agents through the lens of temporal logic formulae, which govern the short-term action plans and long-term goals of the parties involved, and then to exploit this (possibly stochastic) interaction structure to proactively infer intent and plan the AS s actions. The advantages of this formulation are that it is (1) principled (complex interactions among decisionmaking agents are systematically represented via temporal logic formulae, possibly learned), (2) amenable to dynamic programming optimization, and (3) general, as it encompasses a full range of interaction scenarios from collaborative to adversarial.The research objectives are:~ Objective 1. Interaction models: Devise interaction structures in the language of formal methods, and embed them in knowledge-based (e.g., MDP) and data-driven (e.g., deep neural networks) decision-making frameworks.~ Objective 2. Decision-making algorithms: Devise exact and approximate algorithms for the real-time synthesis ofinformation- and game-theoretic action plans that infer intents and anticipate responses.~ Objective 3. Validation: Validate models and algorithms on a state-of-the-art driving simulator, and on a quadrotor test bed whereby a human ~fences~ witha quadrotor.My ultimate project vision is that AS will operate in shared environments by proactively interacting with other agents. As such, this project is closely aligned with the ONR Code 31 program ~Machine Learning, Reasoning and Intelligence,~ and, more in general, with the U.S. Navy s vision of autonomous, adaptive systems that can safely operate in uncertain and unstructured environments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2017
- Source ID
- N000141712433
Entities
People
- Marco Pavone
Organizations
- Office of Naval Research
- Stanford University
- United States Navy