Quartz 5

Aligning Robot and Human Representations

Properties2
descriptionA framework for addressing representation misalignment in robotics, where robot learned representations fail to capture human values, with mathematical formalization and analysis of existing methods.
tagsclippings, robotics, research, ai

1 min read

Aligning Robot and Human Representations

Authors: Andreea Bobu, Andi Peng, Pulkit Agrawal, Julie Shah, Anca D. Dragan

Published: February 3, 2023 (revised January 28, 2024)

DOI: https://doi.org/10.48550/arXiv.2302.01928

Abstract

Current learning approaches suffer from representation misalignment, where the robot’s learned representation does not capture the human’s representation. This paper addresses the critical challenge of ensuring that robot representations align with human values. The authors argue that representation alignment should be explicitly prioritized alongside task learning, propose a mathematical framework for the problem, and contextualize existing methods within this new perspective.

Key Focus Areas

  • Problem definition for representation misalignment in robotic systems
  • Mathematical formulation of alignment objectives
  • Analysis of current representation learning approaches in robotics
  • Future research directions for addressing alignment challenges

Document: 14 pages, 3 figures, 1 table

Subjects: Robotics (cs.RO), Artificial Intelligence (cs.AI), Machine Learning (cs.LG)

Graph View

Backlinks