Talks & Presentations

Risk-Biased Trajectory Forecasting for Safe Human-Robot Interaction

June 27, 2022

Workshop presentation, Risk Aware Decision Making: From Optimal Control to Reinforcement Learning (RSS 2022 Workshop), New York City, NY, USA

Safe planning is critical for human-robot interaction where robot error can cause injury. Risk measures (e.g., CVaR) promote safety of robot plans, whose evaluation is possible in principle using probabilistic forecasts of human motion. Unfortunately, existing generative forecasting methods are unreliable for risk-evaluation due to inaccurate characterization of uncertainty and finite-sampling, possibly leading to overconfident robot behavior. This paper proposes a novel and deliberate biasing of the forecasting model so that any risk evaluation simplifies to a small finite-sampling and expected cost computation. This benefits online planning where only a small number of samples can be taken due to real-time constraints.

Online Trajectory Planning Algorithms for Robotic Systems under Uncertainty in Interactive Environments

July 22, 2021

Doctoral thesis defense, Department of Aeronautics and Astronautics, Stanford University, Stanford, CA, USA

Recent advances in perception, planning, and control have enabled mobile robots to perform complex human-level tasks in many industrial domains. However, most of the robots still lack the capability to consider and address uncertainty, which demands that they be caged or confined to a dedicated, structured workspace. To enable reliable autonomy for “cage-free” robotic operations, this thesis will present computationally efficient algorithms for trajectory planning that can collectively overcome environmental and dynamic uncertainties existing in open, interactive environments. Our approach leverages probability theory and optionally machine learning to quantify the amount of present and future uncertainty. Based on the quantification, we develop online methods for model-based planning and control that either mitigate, avoid the risk of, or are robust against uncertainty to successfully accomplish a given task. Relevant tasks include active perception, Bayesian reinforcement learning for mobile manipulation, and safe autonomous navigation in human crowds. The methods presented in this thesis will ultimately allow intelligent mobile robots to operate in considerably more uncertain and dynamic workspaces than the current industrial standard. This will open up possibilities for various practical applications, including autonomous field robots for persistent environmental monitoring, fully automated driving on urban roads, and autonomous drone flights in densely populated areas for logistics services.

Information-Theoretic Approaches to Active Sensing: Theory and Practice

June 22, 2019

Invited talk, 2nd Workshop on Planning and Adaptive Sampling (RSS 2019 Workshop), Freiburg, Germany

Many mobile robots today are capable of collecting information in the form of sensor readings. However, what sensor readings lead to “desirable” information and how to collect them are both ambiguous and challenging questions; such information could be on parameters of an unknown vector field, or hidden intent of traffic participants on a public road. Information-theoretic active sensing, in which a robot closes the loop from perception to action in search of optimal control under an information-theoretic cost, sheds light on those questions from an optimization perspective. The key components in the information-theoretic active sensing is the online inference of partially observable stochastic processes and the effective planning to reduce their uncertainty. In this talk, I will summarize recent and ongoing projects on this topic within the Multi-Robot Systems Lab at Stanford University. The algorithmic strategies range from RRT path planning to Belief MDP, and the applications include environmental monitoring, target tracking, and intent inference.