FURI | Fall 2025

Risk-sensitive Forward Modeling for Delayed Teleoperation Using Cumulative Prospect Theory

Remote teleoperation of robots faces significant challenges when communication delays prevent operators from receiving real-time feedback, particularly in high-stakes environments like lunar missions where delays can exceed 1.25 seconds. Current teleoperation systems fail to account for human risk perception biases, leading to suboptimal decision-making under uncertainty. This research develops a risk-sensitive forward model that predicts human operator actions by incorporating cumulative prospect theory (CPT), which models how humans actually perceive and respond to risk rather than assuming rational decision-making.

We implement a continuous 2D navigation environment with obstacles and compare multiple machine learning approaches, including variational autoencoders (VAE), Bayesian neural networks, Transformers, and traditional autoencoders for predicting next actions based on current state-action-goal tuples. The forward model integrates CPT-based risk profiles to blend predicted actions with human behavioral biases, accounting for loss aversion and probability weighting that characterize real human decision-making under uncertainty.

Our systematic evaluation using cross-validated hyperparameter tuning demonstrates that different architectures excel in different aspects: VAE models show strong uncertainty quantification capabilities essential for risk assessment, while Transformer models capture complex temporal dependencies in action sequences. The risk-sensitive forward model aims to reduce collision rates and improve mission success in delayed teleoperation scenarios by anticipating human risk-averse behaviors. This work contributes to safer autonomous systems that can predict and adapt to human decision-making patterns, with applications extending beyond teleoperation to human-robot collaboration in manufacturing, healthcare, and emergency response.