Fig. 1. Andrea, Dec 2018. Other portraits by or with robots and roboticists.




PGP key

Address:

Sonneggstrasse 3
ML K 32.2
8092 Zürich
Switzerland

Andrea Censi

I work in the areas of systems and robotics (embodied artificial intelligence).

This is what I do:

I obtained a Ph.D. in Control & Dynamical Systems from 🌴 Caltech 🌴 in 2012 in Richard Murray's group.

I grew up in Italy and obtained an M.Eng. degree in control & robotics from Sapienza University.

👤 CV

Last update: Dec 2020

Supervision

Ph.D. students I work with:

  • Ezzat Elokda* (karma theories, fairness for humans and machines)
  • Dejan Milojevic** (task-driven sensor selection)
  • Marco Wiedner (contextual HMI), industrial PhD with Porsche
  • Alessandro Zanardi* (monadic game theory)
  • Gioele Zardini (co-design, mobility)
  • Julian Zilly (plasticity in deep learning)

Postdocs I work with:

*: co-supervised with Saverio Bolognani
**: co-supervised with EMPA.

Misc

  • Consider attending the (now virtual) Autonomy Talks at ETH Zürich (recordings) to get an overview of research in autonomy / automation at ETH and elsewere.

Teaching

for everybody @ ETH Zürich
Embodied Autonomy The free online Self-driving cars with Duckietown MOOC with EdX will start February 2021. The Autonomous Mobility on Demand (AMOD) class is offered every fall. Check the catalogue for more information.
Applied Category Theory,
co-design
Applied Compositional Thinking for Engineers (ACT4E) is a free online course happening in January 2021. The (more intense) ETH Zürich version of ACT4E will be offered Spring 2021. Check the catalogue for more information.

What's new

2020
Dec 2020 - ACT4E

We just announced a new free online course during January called Applied Compositional Thinking for Engineers (ACT4E). An ETH version for ETH students will come in the Spring.

Nov 2020 - Duckietown becomes a MOOC
The Self-driving cars with Duckietown MOOC on EdX will start in February 2021!
Nov 2020 - AI-DO 5

The AI Driving Olympics edition number 5 is going to be at NeurIPS 2020! See the introduction video below, or take a look at the Challenges server.

Oct 2020 - IROS 2020 workshop on the state of AVs

At the IROS 2020 Workshop on Benchmarking Progress in Autonomous Driving we had very cool debates about the state of the field.

Part 1, Part 2, Part 3, Part 4.

2019

Lots of things happened, but there was no time to write it all down!

May 2019 - AI-DO at ICRA 20219
2018

Lots of things happened, but there was no time to write it all down!

Oct 2018
Dream job: Robotics
2017
2016
  • Jun 2016: During Spring 2016 I taught MIT 2.166 "Autonomous Vehicles", affectionately known as Duckietown, a graduate-level class for makers && thinkers. The elves prepared the robots for us:
  • Apr 2016: I edited the ICRA trailer(s), in 4 episodes.
  • 2015
    • Colleagues roboticists: Please come to the workshop on "The Big Questions in Robotics" at RSS 2015 (Thursday Jul 16).
      (Yes, really, the name is "The Big Questions in Robotics". After our friends in Freiburg organized the workshop "What Sucks in Robotics" we figured that everything goes.)

    Bio news

    • As of 2020, I am working on my own startup (Züpermind), currently in stealth mode.
    • As of July 2019, I am not employed anymore with Aptiv. Between 2016 and 2019 I worked for nuTonomy, an MIT self-driving cars startup. Since 2018 I was Director of Research at Aptiv, the company that acquired nuTonomy. The division has since been spun-off as Motional.
    • As of January 2017, I left MIT to join ETH Zürich.

    Other cool stuff

    The Duckumentary of the first edition of the Duckietown class.

    Previous work

    Click here to see a summary of my research as of 2016. Nowadays the new content is on students' pages and this part is not updated anymore.

    My research

    Here's what's up: The robots are coming! Please see here to check the likelihood that your job will be replaced by a machine. But don't worry too much; the thing is, we don't really know how to design complex robotic systems, despite the headlines.

    I work on the Science of Embodied Autonomy. The engineering applications of my work are towards making complex autonomous robotic systems more robust, more efficient, and easier to design. More generally, I want to understand what are the principles underlying embodied intelligence, both natural and artificial.

    1. Co-Design

    I am working on a new theory of “co-design” to unify all aspects of the design of robotic systems, including energetics, actuation, sensing, and computation.

    For more information, please see the site co-design.science.

    New: an online demo at demo.co-design.science.

    Andrea Censi. A mathematical theory of co-design. Technical Report, Laboratory for Information and Decision Systems, MIT, September 2016. Submitted and conditionally accepted to IEEE Transactions on Robotics. pdf supp. material

    bibtex

    This paper introduces a theory of co-design that describes "design problems", defined as tuples of "functionality space", "implementation space", and "resources space", together with a feasibility relation that relates the three spaces. Design problems can be interconnected together to create "co-design problems", which describe possibly recursive co-design constraints among subsystems.

    read more...

    Andrea Censi. Uncertainty in monotone co-design problems. IEEE Robotics and Automation Letters, February 2017. pdf supp. material

    bibtex

    This paper concerns the introduction of uncertainty in the MCDP framework. Uncertainty has two roles: first, it allows to deal with limited knowledge in the models; second, it also can be used to generate consistent relaxations of a problem, as the computation requirements can be lowered should the user accept some uncertainty in the answer.

    read more...

    2. Robotics education and outreach

    Liam Paull, Jacopo Tani, Heejin Ahn, Javier Alonso-Mora, Luca Carlone, Michal Cap, Yu Fan Chen, Changhyun Choi, Jeff Dusek, Daniel Hoehener, Shih-Yuan Liu, Michael Novitzky, Igor Franzoni Okuyama, Jason Pazis, Guy Rosman, Valerio Varricchio, Hsueh-Cheng Wang, Dmitry Yershov, Hang Zhao, Michael Benjamin, Christopher Carr, Maria Zuber, Sertac Karaman, Emilio Frazzoli, Domitilla Del Vecchio, Daniela Rus, Jonathan How, John Leonard, and Andrea Censi. Duckietown: an open and inexpensive and flexible platform for autonomy education and research. In IEEE International Conference on Robotics and Automation (ICRA). Singapore, May 2017. pdf supp. material

    bibtex

    This paper describes the Duckiebot and its software. With 29 authors, we made the record for a robotics conference.

    read more...

    Jacopo Tani, Liam Paull, Maria Zuber, Daniela Rus, Jonathan How, John Leonard, and Andrea Censi. Duckietown: an innovative way to teach autonomy. In EduRobotics 2016. Athens, Greece, December 2016. pdf supp. material

    bibtex

    This paper describes the course design for Duckietown: learning objectives, teaching methods, etc.

    read more...

    3. Neuromorphic / bio-inspired control

    I'm interested in co-design problems that couple sensing, computation, and actuation in a non-trivial way, especially from the point of view of design minimality and "joint inference and control".

    In particular, I'm interested in the robotics applications of event-based neuromorphic vision sensors. These are a new kind of sensor that outputs a low-latency stream of eevents, generated every time there is a change in the local brightness perceived by a pixel, rather than a periodic series of frames.

    Events from a neuromorphic sensor, in real time (left) and slowed down 50x (center). On the right, events superimposed with a CMOS sensor's output.

    My research is guided by the questions:

    • Theory: How can we formally say that this sensor class is better than another? For what tasks? In what environments?
    • Algorithms: How can we obtain "zero-latency" event-based controllers?
    • Systems: How can we integrate these sensors in existing control architectures?

    Andrea Censi. Efficient Neuromorphic Optomotor Heading Regulation. In American Control Conference (ACC). Chicago, IL, July 2015. doi

    bibtex

    Andrea Censi, Erich Mueller, Emilio Frazzoli, and Stefano Soatto. A Power-Performance Approach to Comparing Sensor Families, with application to comparing neuromorphic to traditional vision sensors. In IEEE International Conference on Robotics and Automation (ICRA). May 2015.

    bibtex

    Andrea Censi and Davide Scaramuzza. Low-latency event-based visual odometry. In IEEE International Conference on Robotics and Automation (ICRA). May 2014. pdf supp. material slides

    bibtex

    Andrea Censi, Jonas Strubel, Christian Brandli, Tobi Delbruck, and Davide Scaramuzza. Low-latency localization by active led markers tracking using a dynamic vision sensor. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 891–898. Tokyo,Japan, November 2013. pdfdoi supp. material slides

    bibtex

    Sawyer B. Fuller, Michael Karpelson, Andrea Censi, Kevin Y. Ma, and Robert J. Wood. Controlling free flight of a robotic fly using an onboard vision sensor inspired by insect ocelli. Journal of the Royal Society Interface, August 2014. video supp. material

    bibtex

    The world is full of natural robots, which are informally called “animals.” I have worked on the identification of fruit-fly stimulus-elicited behavior, with Dickinson and Straw. This kind of work is interesting for an engineer, because it can be seen as reverse-engineering of an existing solution. Lately, I'm thinking about how to make this duality between analysis (biology) and synthesis (robotics) more formal. I'm very interested in possible collaborations on this topic.

    Andrea Censi*, Andrew D. Straw*, Rosalyn W. Sayaman, Richard M. Murray, and Michael H. Dickinson. Discriminating external and internal causes for saccade initiation in freely flying Drosophila. PLOS Computational Biology, February 2013. pdfdoi supp. material slides

    bibtex

    Some recent workshops on these topics:

    4. Sensorimotor Learning and Bootstrapping

    Imagine you are a brain-in-a-vat that wakes up connected to an unknown body through two streams of uninterpreted observationd and commands, without any prior knowledge of the sensors, actuators or environment. Would you be able to learn a model of your body and use it to perform useful tasks? This is the "bootstrapping scenario": the learning problem for an embodied agent in the limit of prior knowledge tending to zero.

    Uninterpreted streams of observations and commands from a robotic sensorimotor cascade.

    My research in this field is guided by the questions:

    • Theory: How much prior knowledge is needed by an agent?
    • Algorithmics: What are tractable classes of models for sensorimotor learning?
    • Systems: How can we introduced learning and adaptivity functionality in traditional robotic control systems?
     

    The video shows learning of a bilinear model of a sensorimotor cascade for a camera. The agent starts with no previous knowledge on the sensor geometry, and by correlating observations with commands, it can learn a generative model for the data. The same model can be used for learning the dynamics of different sensors (range-finder, camera, field sampler). See many other videos of related experiments.

    Representative works:

    Andrea Censi and Richard M. Murray. Bootstrapping bilinear models of Simple Vehicles. International Journal of Robotics Research, 34:1087–1113, July 2015. pdfdoi video supp. material slides

    bibtex

    Andrea Censi and Davide Scaramuzza. Calibration by correlation using metric embedding from non-metric similarities. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35:2357–2370, 10 2013. pdfdoi supp. material

    bibtex

    Andrea Censi, Adam Nilsson, and Richard M. Murray. Motion planning in observations space with learned diffeomorphism models. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2860–2867. Karlsruhe, Germany, 5 2013. pdfdoi supp. material

    bibtex

    Andrea Censi and Richard M. Murray. Learning diffeomorphism models of robotic sensorimotor cascades. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA). Saint Paul, MN, May 2012. pdfdoi supp. material slides

    bibtex

    Andrea Censi and Richard M. Murray. Uncertain semantics, representation nuisances, and necessary invariance properties of bootstrapping agents. In Joint IEEE International Conference on Development and Learning and Epigenetic Robotics. Frankfurt, Germany, August 2011. pdfdoi slides

    bibtex

    Andrea Censi, Antonio Franchi, Luca Marchionni, and Giuseppe Oriolo. Simultaneous calibration of odometry and sensor parameters for mobile robots. IEEE Transactions on Robotics, 29(2):475–492, April 2013. pdfdoi supp. material

    bibtex

    My dissertation:

    Andrea Censi. Bootstrapping Vehicles: a formal approach to unsupervised sensorimotor learning based on invariance. Technical Report, California Institute of Technology, 2012. pdf supp. material slides

    bibtex

    Recent workshop:

    5. “Classic” robotics perception

    At the beginning of my research career I worked on what we can now call “classic” robotics perception and planning problems.

    Some recent work on the problem of pose-graph optimization, with Luca Carlone:

    Luca Carlone and Andrea Censi. From Angular Manifolds to the Integer Lattice: Guaranteed Orientation Estimation with Application to Pose Graph Optimization. IEEE Transactions on Robotics, April 2014. pdfdoi supp. material slides

    bibtex

    Luca Carlone, Andrea Censi, and Frank Dellaert. Coherent measurements selection via l1 relaxation: an approach to robust estimation over graphs. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). October 2014.

    bibtex

    A paper with Davide Scaramuzza on visual odometry:

    Davide Scaramuzza, Andrea Censi, and Kostas Daniilidis. Exploiting motion priors in visual odometry for vehicle-mounted cameras with non-holonomic constraints. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). San Francisco, CA, September 2011. pdfdoi

    bibtex

    Some papers on the algorithmics and the accuracy of pose tracking and localization using range data:

    Andrea Censi. On achievable accuracy for pose tracking. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA). Kobe, Japan, May 2009. pdfdoi supp. material slides

    bibtex

    Andrea Censi. An ICP variant using a point-to-line metric. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA). Pasadena, CA, May 2008. pdfdoi supp. material slides

    bibtex

    Andrea Censi. An accurate closed-form estimate of ICP's covariance. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 3167–3172. Rome, Italy, April 2007. pdfdoi supp. material slides

    bibtex

    Andrea Censi. On achievable accuracy for range-finder localization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 4170–4175. Rome, Italy, April 2007. pdfdoi supp. material slides

    bibtex

    Point-to-point ICP (left) vs point-to-line ICP (right).

    Software

    Most of my papers come with software and datasets in the spirit of reproducible research (subject to time constraints...). These are the software packages that were polished and documented enough for wider use.

    Software users: It's always great to know that my software is used for something cool. Please send me an email if you do.

    Some general-purpose software that came out as a side-effect of my research:

    • PyContracts is a Python package that allows to declare constraints on function parameters and return values. It supports a basic type system, variables binding, arithmetic constraints, and has several specialized contracts (notably for Numpy arrays). As a quick intro, please see this presentation about PyContracts.
    • Compmake is a Python library that provides “Make”–like facilities to a Python application, including job management and parallelization (multiple CPU on a single host, cluster computing using SGE, and experimental support for cloud computing using Multyvac).

    Some robotics-specific software packages:

    • The CSM scan matcher. This is currently maintained by Kuka. It is integrated in many ROS packages.

      Andrea Censi. An ICP variant using a point-to-line metric. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA). Pasadena, CA, May 2008. pdfdoi supp. material slides

      bibtex
    • This odometry + sensor calibration package corresponding to the following paper.

      Andrea Censi, Antonio Franchi, Luca Marchionni, and Giuseppe Oriolo. Simultaneous calibration of odometry and sensor parameters for mobile robots. IEEE Transactions on Robotics, 29(2):475–492, April 2013. pdfdoi supp. material

      bibtex

    Professional Service

    Personal

    The ICRA 2015 trailer