Learning and Leveraging Conventions in the Design of an Adaptive Haptic Shared Control for Steering a Semi-Automated Vehicle

Vahid Izadi and Amirhossein Ghasemi

Email: ah.Ghasemi@uncc.edu

Website: https://coefs.uncc.edu/aghasem1

​

​

​

Spring 2022

Motivation

Autonomous Ground Vehicles hold great potential for both military and commercial applications

​

2

Faults Happen

3

When is Semi-Autonomy is Helpful?

4

Detect-Detect

Miss-Detect

Detect-Miss

Miss-Miss

5

Current Transfer Control Approaches

Input Mixing Shared Control:

Shortcomings: No Direct Connection between Human and Automation

Haptic Shared Control: referred to as physical human-robot interaction (pHRI), is a sharing scheme that involves physical coupling between a human partner and co-robot.

Input Switching Control:

Shortcomings:

• misinterpretation or misappropriation of responsibility (called mode errors)
• incomplete understanding of the environment state (loss of situation awareness)

Haptic Shared Control

6

Human Driver

Automation System

Adding Automation Can Cause Conflicts

7

Detect-Detect

Miss-Detect

Detect-Miss

Miss-Miss

Research Goals

Intention (Right or Left) and Role Negotiations (leader/follower) can lead to successful collaboration.

Conventions can narrow this to a subset of these equilibria (or norms) to which the team might more naturally gravitate. Current control transfer mechanisms are built based on fixed conventions.

Goal: Develop the principles of convention formation in a haptic shared control framework to determine optimal handover strategies in uncertain circumstances.

Challenge: In a multi-agent repeated game context, there might be multiple strategies for solving a problem (several equilibria), with some preferable than others.

Goal 1: Create a modular structure that separates partner-specific conventions from task-dependent representations.

​

Goal 2: Characterize the map from the space of conventions to outcomes in driver-automation collaboration

​

Goal 3: Develop Techniques to influence and guide the agents (automation and driver) to reach a desirable shared convention.

Research Goals

Goal 1: Creating a modular structure that separates partner-specific conventions from task-dependent representations.

Let’s define a cost function for each agent

Goal 2: Creating a map from the space of conventions to outcomes in multiagent and driver-automation collaboration

Nash strategy

Zero-sum vs. Non-zero Sum Games

Overall Goal: Minimize the combined cost

Challenge: favor one more than the other

Cooperative-Competitive Game

This co-co concept models a situation where one agent pays/receives an incentive to implement a strategy that minimizes the combined cost function. The co-co game requires communication between the agents.

14

Examples of Convention Surfaces

Active Safety

Neutral

Assistive Behavior

Cooperative Cost Value

Competitive Cost Value

15

Three sample convention-based paradigms

Neutral Behavior

Selfless Behavior

Selfish Behavior

---- Automation’s preferred path

…..Human’s preferred path

____Vehicle path

16

Goal 3: Develop techniques to influence and guide the agents to reach shared conventions.

​

​

RL-MPC for Policy Search

17

High-Level Decision Variables

Reinforcement Learning for High-Level Policy

Haptic Shared Control

MPCs

Dynamics

Actions

References

States

Observations

Automation’s preferred path

Human’s preferred path

Vehicle path

Differential Torque (fight)

Human’s weight for being selfish

Automation’s weight for being selfish

Cost Values for the Cooperative Part

Cost Values for the Competitive Part

Future Works

19

Thank You!