Reproduce “Probabilistically Safe Robot Planning with Confidence-Based Human Predictions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jaime F. Fisac, Andrea Bajcsy, Sylvia L. Herbert, David Fridovich-Keil, Steven Wang, Claire J. Tomlin, Anca D. Dragan
The paper presents a robot planner that treats how well a human-motion model explains observed actions as a changing confidence signal. It updates a Bayesian belief over β and combines the resulting predictions with FaSTrack’s worst-case tracking bound. In simulations replaying 48 trajectories from 16 people, adaptive confidence improved safety proxy and travel time in unmodeled-obstacle cases, but not universally.
Key insights
- Reinterpreting the Boltzmann rationality coefficient (β) as the robot's real-time confidence in its human motion model provides a principled way for robots to assess the reliability of their predictions.
- Dynamically adapting robot conservativeness based on online model confidence leads to demonstrably safer and more efficient trajectories than static, fixed-confidence planning.
- Combining worst-case robot tracking error guarantees with time-varying probabilistic human predictions yields a quantitative probabilistic safety certificate, enhancing rigor in safety assurance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.