[Paper Review] Personalized Dynamics Models for Adaptive Assistive Navigation Interfaces.

Eshed Ohn-Bar, Kris Kitani+1 more|arXiv (Cornell University)|Apr 11, 2018

Tactile and Sensory Interactions81 references3 citations

TL;DR

This paper proposes PING, a personalized instruction generation agent that uses model-based reinforcement learning with a novel weighted majority regression algorithm to rapidly adapt navigational guidance for visually impaired users. By learning individualized dynamics models, PING improves long-horizon position prediction by over 1 meter on average (20-second horizon) and enhances performance at critical navigation junctures like turns.

ABSTRACT

We explore the role of personalization for assistive navigational systems (e.g., service robot, wearable system or smartphone app) that guide visually impaired users through speech, sound and haptic-based instructional guidance. Based on our analysis of real-world users, we show that the dynamics of blind users cannot be accounted for by a single universal model but instead must be learned on an individual basis. To learn personalized instructional interfaces, we propose PING (Personalized INstruction Generation agent), a model-based reinforcement learning framework which aims to quickly adapt its state transition dynamics model to match the reactions of the user using a novel end-to-end learned weighted majority-based regression algorithm. In our experiments, we show that PING learns dynamics models significantly faster compared to baseline transfer learning approaches on real-world data. We find that through better reasoning over personal mobility nuances, interaction with surrounding obstacles, and the current navigation task, PING is able to improve the performance of instructional assistive navigation at the most crucial junctions such as turns or veering paths. To enable sufficient planning time over user responses, we emphasize prediction of human motion for long horizons. Specifically, the learned dynamics models are shown to consistently improve long-term position prediction by over 1 meter on average (nearly the width of a hallway) compared to baseline approaches even when considering a prediction horizon of 20 seconds into the future.

Motivation & Objective

• To address the limitations of universal models in capturing individual differences in blind users' mobility behaviors.
• To develop a personalized instructional interface that adapts quickly to individual user responses and navigation patterns.
• To improve long-horizon human motion prediction for planning effective assistive navigation guidance.
• To enhance performance at critical navigation points such as turns and veering paths through user-specific modeling.
• To enable real-time adaptation by learning dynamics models end-to-end using user feedback.

Proposed method

• PING employs a model-based reinforcement learning framework to learn personalized state transition dynamics for each user.
• It uses an end-to-end learned weighted majority-based regression algorithm to adapt the dynamics model to individual user reactions.
• The framework incorporates user-specific mobility nuances, obstacle interactions, and current navigation tasks into the dynamics modeling process.
• It predicts human motion over long horizons (up to 20 seconds) to ensure sufficient planning time for guidance.
• The model is trained on real-world user data to capture individual variations in response patterns and navigation behavior.
• Personalized dynamics models are continuously refined using feedback from user interactions during navigation.

Experimental results

Research questions

• RQ1 Can personalized dynamics models significantly improve long-term position prediction in assistive navigation for visually impaired users?
• RQ2 How does PING’s end-to-end weighted majority regression compare to baseline transfer learning in adapting to individual users?
• RQ3 To what extent does personalization enhance performance at critical navigation junctures like turns and veering paths?
• RQ4 Can learned dynamics models improve planning time by enabling accurate long-horizon motion prediction?
• RQ5 How do individual mobility nuances and obstacle interactions affect the performance of personalized navigation systems?

Key findings

• PING learns personalized dynamics models significantly faster than baseline transfer learning approaches on real-world data.
• The learned dynamics models improve long-term position prediction by over 1 meter on average across a 20-second prediction horizon.
• Performance gains are most pronounced at critical navigation points such as turns and veering paths.
• Personalization enables better reasoning over user-specific mobility patterns and obstacle interactions.
• The framework consistently outperforms non-personalized models in predicting future user positions over extended timeframes.
• The end-to-end weighted majority regression algorithm enables rapid adaptation to individual user response dynamics.