Vincent Pacelli


I am a PhD candidate at Princeton University advised by Anirudha Majumdar as part of the Intelligent Robot Motion IRoM (IRoM) Lab.

My work is on understanding necessary and sufficient amounts of task-relevant information needed to build a performant robot, both theoretically and experimentally.

Broadly, my interests are in topics involving dynamics and control, optimization, and connections between problems in robotics and natural sciences.


Current Research

My research to date has focused on studying the role of task-relevant information in control problems and developing controllers that exploit it.

Task-Relevancy and Bounded Rationality Robots

Bounded rationality is the concept that agents (e.g., humans and robots) only have limited computational and sensing resources available to them, so truely rational decision-making is infesible. Bounded rationality strategies focus on only processing the most relevant information for a given task, leading to efficient and robust decision-making.

A canonical example is the gaze heuristic, which is a simple feedback law humans employ to catch a ball. Specifically, humans follow a trajectory that keeps the ball at a fixed point in their visual field. This strategy is provably correct, computationally simple, and adapts robustly to disturbances such as wind gusts or visual background noise as long as the task-relevant information, i.e. the angle between the ball and the eye line, is detectable. The gaze heuristic is a categorically different strategy than the generic recipies for motion planning employed in robotics, where state estimation and model-preditive control are tightly coupled. This method requires significant computation to process rich sensor information and the tight coupling of estimation and control lacks robustness without rapidly replanning.

My work in this area has focused on algorithmically identifying the (often low-dimensional) task-relevant variables in optimal control problems, designing controllers that depend on these variables only, and demonstrating the performance-robustness tradeoffs of these controllers. These goals are achieved through an approach using overlapping tools from fields such as, information theory, statistical mechanics, and differential privacy.

Fundamental Limits in Robotics

Some of my recent work has been on understanding fundamental limits on the performance achievable by a robot equipped with different sensors on an optimal control problem. The aim is to answer questions such as: does an autonomous car require different sensing modalities, or is a vision system enough? In principle, if both systems capture the relevant information for the task, then performance is not gained by adding more sensors to the vehicle. My work makes early steps toward answering this question by deriving estimatable lower-bounds on control cost in terms of the information a sensor may provide using generalized versions of Fano’s Inequality.

Future Directions

I have a growing interest in the interplay between the field of robotics and the natural sciences. Generally, these fields seek answers to different questions, leading to distinctly motivated problems, unique tools for solving them, and lexical barriers between communities. However, robotics is a nacent field still searching for the proper modeling formulisims and an understanding of the limits of what can be reasonably expected from a robot. In comparison to robotics, the natural sciences excel in these areas, and there is untapped potential in importing useful mathematical formalisms from them into robotics. Occasionally, successful cross-polination has occured, for example path integral methods methods leading to new optimal control algorithms in robotics, but there is still much work to be done in bridging the gap between robotics and the natural sciences.


Publications

  1. "Fundamental Performance Limits for Sensor-Based Robot Control and Policy Learning."
    A. Majumdar and V. Pacelli. RSS 2022. [arxiv] [pub]
  2. "Robust Control Under Uncertainty via Bounded Rationality and Differential Privacy."
    V. Pacelli and A. Majumdar. ICRA 2022. [arxiv] [pub]
  3. "Invariant Policy Optimization: Towards Stronger Generalization in Reinforcement Learning."
    A. Sonar, V. Pacelli, and A. Majumdar. L4DC 2021. [arxiv] [pub]
  4. "Learning Task-Driven Control Policies via Information Bottlenecks."
    V. Pacelli and A. Majumdar. RSS 2020. [arxiv] [pub]
  5. "Task-Driven Estimation and Control via Information Bottlenecks."
    V. Pacelli and A. Majumdar. ICRA 2019. [arxiv] [pub]
  6. "Integration of Local Geometry and Metric Information in Sampling-Based Motion Planning."
    V. Pacelli, O. Arslan, and D. E. Koditschek. ICRA 2018. [pub]
  7. "Sensory Steering for Sampling-Based Motion Planning."
    O. Arslan, V. Pacelli, and D. E. Koditschek. IROS 2017. [pub]

Talks

  1. "Robust Control for Robots via Minimal-Information Policies."
    V. Pacelli and A. Majumdar. APS March Meeting 2021.
  2. "Task-Driven Representations for Robot Estimation and Control."
    V. Pacelli and A. Majumdar. NERC 2019.