# Probabilistically Safe Robot Planning with Confidence-Based Human Predictions

**Authors:** Jaime F. Fisac, Andrea Bajcsy, Sylvia L. Herbert, David Fridovich-Keil, Steven Wang, Claire J. Tomlin, Anca D. Dragan  
**arXiv:** [1806.00109](https://arxiv.org/abs/1806.00109)  
**Date:** 2018-05-31  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1806.00109v1/

## Key points

- The paper addresses the challenges of safely operating robots around humans
- Predictive models of human motion can be employed to ensure safety
- Simplifying assumptions in these models can lead to degraded predictions and pose a risk to safety
- The authors propose a real-time Bayesian framework that allows robots to modulate their predictions based on the model's confidence in predicting human behavior
- The approach combines worst-case tracking error guarantees for physical robots with probabilistic time-varying human predictions to generate assured autonomous robot motion and yield a quantitative, probabilistic safety certificate
- The approach is demonstrated using a quadcopter navigating around a human and highlights that rational human actions under a particular model can serve as an indicator of its ability to describe current motion and improve prediction accuracy while ensuring safety.

## Summary

The paper titled "Probabilistically Safe Robot Planning with Confidence-Based Human Predictions" addresses the challenges of safely operating robots around humans. To ensure safety, predictive models of human motion can be employed; however, these models often introduce simplifying assumptions that cannot capture the full complexity of human behavior. This can lead to degraded predictions when observed human behavior deviates from the assumed structure and poses a risk to safety. To address this issue, the authors propose a real-time Bayesian framework that allows robots to modulate their predictions based on the model's confidence in predicting human behavior. By leveraging recent work in provably-safe trajectory planning and combining worst-case tracking error guarantees for physical robots with probabilistic time-varying human predictions, the authors generate assured autonomous robot motion and yield a quantitative, probabilistic safety certificate. The approach is demonstrated using a quadcopter navigating around a human and highlights that rational human actions under a particular model can serve as an indicator of its ability to describe current motion and improve prediction accuracy while ensuring safety.