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 dissertation work is on understanding the 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 dynamics and control, information theory, optimization, and connections between problems in robotics and natural sciences.


Research Overview

My research to date has focused on studying the role of task-relevant information in control problems, the theoretical benefits of making control decisions using only task-relevant information, algorithmically designing these controllers, and how much task-relevant information a sensor must provide to achieve performant control.

Task-Relevancy and Bounded Rationality Robots

Today, robots are equipped with powerful, high-dimensional, general-purpose sensors (e.g., cameras), which allow general purpose control algorithms to achieve high performance on many challenging tasks. However, the generic nature of these algorithms requires that large amounts of sensor information is processed regardless of whether or not it is relevant to the robot’s task, and disturbances irrelevant components of the sensor observation can significantly degrade the performance of the closed-loop system.

In contrast, humans are equipped with powerful, high-dimensional, general-purpose, sensors (e.g., our eyes) but they rely on task-relevant heuristics to perform dexterous tasks, like ball-catching, robustly by monitoring only a few task-relevant variables that capture most of important information. This kind of limited, but effective decision-making is known as employing bounded rationality. 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 in conjunction with experiments in simulation and on real hardware platforms.

Fundamental Limits in Sensing and Control for 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 the cost of sensing and processing additional information does not yield a meaningful increase in performance.

Establishing such limits is largely an open problem in robotics. 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 emerging techniques in information theory (e.g., new generalizations of Fano’s inequality) and non-equilibrium thermodynamics.

Future Interests

I have a growing interest in the interplay between fundamental problems in robotics and the natural sciences. Problems in natural sciences lead to the development of many useful theoretical tools and modeling formalisms for describing fundamental problems in these fields. Robotics, as a comparatively young discipline formed from the intersection between engineering, computer science, and other fields, is still searching for the right formalism to describe its unique problems. Adapting tools from the natural science broaden the vocabulary of the field and deepen our understanding of robotics.


Publications

  1. Robust Control Under Uncertainty via Bounded Rationality and Differential Privacy (Extension).
    V. Pacelli and A. Majumdar. IJRR 2022 (In Preparation).
  2. Fundamental Performance Limits for Sensor-Based Robot Control and Policy Learning.
    A. Majumdar and V. Pacelli. RSS 2022. [arxiv] [pub]
  3. Robust Control Under Uncertainty via Bounded Rationality and Differential Privacy.
    V. Pacelli and A. Majumdar. ICRA 2022. [arxiv] [pub]
  4. Invariant Policy Optimization: Towards Stronger Generalization in Reinforcement Learning.
    A. Sonar, V. Pacelli, and A. Majumdar. L4DC 2021. [arxiv] [pub]
  5. Learning Task-Driven Control Policies via Information Bottlenecks.
    V. Pacelli and A. Majumdar. RSS 2020. [arxiv] [pub]
  6. Task-Driven Estimation and Control via Information Bottlenecks.
    V. Pacelli and A. Majumdar. ICRA 2019. [arxiv] [pub]
  7. Integration of Local Geometry and Metric Information in Sampling-Based Motion Planning.
    V. Pacelli, O. Arslan, and D. E. Koditschek. ICRA 2018. [pub]
  8. 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.