Andrea Censi's website and blog updates

Recent work on diffeomorphism learning

andrea — Tue, 16 Apr 2013 00:08:51 +0000

Some recent work on diffeomorphisms learning: Accurate recursive learning of uncertain diffeomorphism dynamics.

Localization with dynamic vision sensor

andrea — Tue, 16 Apr 2013 00:07:29 +0000

Recent work with the Zurich folks: Low-latency localization by active led markers tracking using a dynamic vision sensor.

Compmake keeps calm and carries on: a fault-tolerant “make” for parallelizable batch Python computations

andrea — Sat, 09 Feb 2013 09:39:14 +0000

A few years ago, at a SLAM summer school in Oxford, I attended a talk by Frank Dellaert about a scientist’s “secret weapons”. I believe he was channeling his advisor Herb Simon in saying that every scientist must have a secret weapon that give them an edge against the others. (For the record, Frank’s secret weapon is OCaml.) In this age where most science is highly computational and data-based, the secret weapons are often software tools. Compmake has become one of my secret weapons.

Compmake is a nonobtrusive module that provides Makefile--like facilities to Python programs, including familiar commands such as make and clean, zero-effort parallelization, caching of results (the program can be interrupted and restarted), a console interface, and many other goodies. I have been relying on it daily since a couple of years and recently I polished it enough that it can be used by others. Here's a quick introduction; see the webpage for the full documentation.

Compmake has been designed primarily for handling long computational-intensive batch processes that can be decomposed in smaller granular jobs. To use Compmake, you have to minimally modify your Python program, such that it can understand the processing layout. Basically, each function call of the kind y = f(x) becomes y = comp(f, x), like the following figure shows. Compmake takes care of the rest. This simple modification is enough to solve most of the frustrating problems that I personally encountered in software development.

To install Compmake, use pip install compmake. You can try the following examples using the demo example.py. One way to use Compmake is to use the compmake executable. If you write:

$ compmake example -c make            # runs serially

the module example will be imported and the jobs defined there using comp will be loaded in the DB. The command make passed with the -c switch executes the jobs serially.

Parallel execution: To run jobs in parallel, use the parmake command:

$ compmake example -c "parmake n=6"   # runs at most 6 in parallel

There are all sorts of configuration options for being nice to other users of the machine; for example, Compmake can be instructed not to start other jobs if the CPU or memory usage is already above a certain percentage:

$ compmake --max_cpu_load=50 --max_mem_load=50 example -c "clean; parmake"

Console: A console is displayed if you just run:

$ compmake example

Some useful commands are ls, make, clean, and parmake, which all do what you think they do. Write help for a list of all commands.

Selective remake: You can selectively remake part of the computations. For example, suppose that you modify the draw() function, and you want to rerun only the last step. You can achieve that by::

$ compmake example -c "remake draw*"

Compmake will reuse part of the computations (func1 and func2) but it will redo the last step.

Tolerance to failures: If some of the jobs fail (e.g., they throw an exception), compmake will go forward with the rest. To see this behavior, download the file example_fail.py, which defines jobs that fail. If you run::

$ compmake example_fail -c "parmake n=4"

you will see how Compmake completes all jobs that can be completed.

Give it a go, and let me know how it goes.

The most useful Numpy function (if you work with tensors)

andrea — Wed, 06 Feb 2013 23:03:55 +0000

I discovered np.einsum, the most useful of all Numpy library functions.

Suppose you want to compute:

$$ T^{svi} = \sum_x \sum_j P^{sx} M^{v}_{xj} Q^{ij} $$

Then you can just describe this operation using indices:

T = np.einsum("sx,vxj,ij -> svi", P, M, Q)

It's more compact than the LaTeX expression!

Another example:

$$ M^{v}_{xj} = \sum_{x} \sum_{j} T^{svi} P^{-1}_{sx} Q^{-1}_{ij} $$

becomes:

    M = np.einsum("svi,sx,ij -> vxj", T, P_inv, Q_inv)

]]>

Pysnip, a package for running large Python jobs from LaTeX

andrea — Wed, 28 Nov 2012 04:03:02 +0000

Pysnip is a small project that I wrote for running Python code from LaTeX. I found this workflow useful for autogenerating tables and reports on the fly, as in Chapter 13 of my dissertation.

Pysnip is based on python.sty, which calls the Python executable directly from latex. That package is good and recommended for small jobs.

Pysnip instead generates external Python files, and these are run (possibly in parallel) using compmake (another project of mine which was essential to survival).

With pysnip you include arbitrary Python code using the \\py command:

\usepackage{pysnip}

\py{ print('hello'); }

and then execute it as follows:

# creates files in snippets/
$ pdflatex file.tex 
# run snippets
$ pysnip-make
# compile final file
$ pdflatex file.tex

New preprint: From Angular Manifolds to the Integer Lattice: Guaranteed Orientation Estimation

andrea — Sat, 24 Nov 2012 13:09:16 +0000

This is some recent work on estimation on manifolds, in collaboration with Luca Carlone.

Luca Carlone and A.C.. From Angular Manifolds to the Integer Lattice: Guaranteed Orientation Estimation with Application to Pose Graph Optimization. ArXiv e-prints, November 2012. pdf supp. material slidesbibtex

@article{carlone12angular_preprint,
        author = "Carlone, Luca and Censi, Andrea",
        title = "{From Angular Manifolds to the Integer Lattice: Guaranteed Orientation Estimation with Application to Pose Graph Optimization}",
        url = "http://www.lucacarlone.com/index.php/resources/research/mole2d",
        journal = "ArXiv e-prints",
        month = "November",
        slides = "http://purl.org/censi/research/2013-mole2d-slides.pdf",
        primaryclass = "cs.RO",
        year = "2012",
        keywords = "Computer Science - Robotics, Mathematics - Optimization and Control, 68T40, I.2.9, G.1.6, G.3, G.4",
        pdf = "http://arxiv.org/pdf/1211.3063v1",
        archiveprefix = "arXiv"
    }

Pose optimization is what is used in SLAM to optimize the map after pose-pose and pose-features correspondences have been established. The variables in this problem are poses living on the nodes of a graph, and measurements are relative measurements along the graph edges. The problem is hard because orientations live on a manifold with nontrivial topology, which makes the problem nonlinear, nonconvex, and with multiple minima. Luca and I try to solve the subproblem of orientation estimation. We find a way to convert the problem to an unconstrained optimization problem on integers. This makes it possible to solve the problem globally and return all likely guesses for the orientation.

Dissertation online

andrea — Sat, 20 Oct 2012 18:36:32 +0000

Version 1.3 of my dissertation is now online.

new NRI project

andrea — Mon, 01 Oct 2012 16:46:20 +0000

The NRI proposal Improved safety and reliability of robotic systems by faults/anomalies detection from uninterpreted signals of computation graphs was funded. Looking forward to buy lots of robots and acquire plenty of data.

The Turing test revisited

andrea — Mon, 01 Oct 2012 09:35:56 +0000

Here’s a revisitation of the Turing test for our modern times.

I wanted to disable the voicemail on my phone, so I used T-Mobile’s tech support chat line. The ordeal took a good half an hour, with minutes passing between each interaction.

My guess is that most of this was automated. For example, after the question “Can I have your PIN please?” I received no answer to my question on how to look it up, but once I wrote “is it possible it’s just 0000?” the answer was immediate. This makes me think of simple pattern matching (in this case, looking for a 4-digit number) in the context of a scripted interaction.

The question is: at which point did I start talking with a human (if ever)?

Please wait while we find an agent to assist you… You are currently at position number 1 in the queue.

You have been connected to !Roselita M.

Andrea Censi: I want to disable my voicemail. How can I do it?

!Roselita M: Thank you for contacting T-Mobile Live Chat, my name is Rose M. and my rep ID is 145-0911. I’ll be more than happy to assist you. Let me pull up your account first.

!Roselita M: Please give me a few moments while I’m waiting for my resources to pull up.

!Roselita M: Andrea, please bear with me. I still need a few more time. Thanks.

!Roselita M: Thank you for patiently waiting. Can I have the make and model of your device?

Andrea Censi: it’s a samsung

Andrea Censi: cheap flip phone

Andrea Censi: not sure which model

!Roselita M: Is it a T-Mobile phone?

Andrea Censi: yes

Andrea Censi: one of the cheapest

!Roselita M: Can I have your PIN please?

Andrea Censi: uhm.. I don’t remember it. Can I look it up on the phone somwhere?

Andrea Censi: is it possible it’s just 0000?

!Roselita M: Thank you.

!Roselita M: May I have your phone’s IMEI? You can find out the 15-digit IMEI to the phone by dialing *#06# on the phone, or checking under the battery.

Andrea Censi: 01219500371524/802

!Roselita M: Thank you. I will check the procedures for your handset.

!Roselita M: Andrea, I still need more time for this.

!Roselita M: Thank you for waiting. I’ve already disabled the voicemail on your account but please note that it can take up to 2 hours for the system to recognize the change. Will there be anything else?

Andrea Censi: no, thanks, that’s all

New preprint: Motion planning in observations space with learned diffeomorphism models.

andrea — Sun, 30 Sep 2012 13:13:07 +0000

Some recent work done on the integration of bootstrapped models with more traditional planning techniques.

A.C., Adam Nilsson, and Richard M. Murray. Motion planning in observations space with learned diffeomorphism models. Technical Report CaltechCDSTR:2012.004, California Institute of Technology, 2012. To appear in ICRA 2013. pdf supp. materialbibtex

@techreport{censi12dptr1_preprint,
        author = "Censi, Andrea and Nilsson, Adam and Murray, Richard M.",
        kind = "preprint",
        title = "Motion planning in observations space with learned diffeomorphism models.",
        url = "http://purl.org/censi/2012/dptr1",
        abstract = "We consider the problem of planning motions in observations space, based on learned models of the dynamics that associate to each action a diffeomorphism of the obser- vations domain. For an arbitrary set of diffeomorphisms, this problem must be formulated as a generic search problem. We adapt established algorithms of the graph search family. In this scenario, node expansion is very costly, as each node in the graph is associated to an uncertain diffeomorphism and corresponding predicted observations. We describe several improvements that ameliorate performance: the introduction of better image similarities to use as heuristics; a method to reduce the number of expanded nodes by preliminarily identifying redundant plans; and a method to pre-compute composite actions that make the search efficient in all directions.",
        number = "CaltechCDSTR:2012.004",
        note = "To appear in ICRA 2013.",
        year = "2012",
        pdf = "http://purl.org/censi/research/2012-dptr1.pdf",
        institution = "California Institute of Technology"
    }

Using learned diffeomorphism models of the dynamics of cameras and range-finders, we formulate motion planning as a planning problem in the observations space. Nodes/states are (uncertain) images; actions/edges are (uncertain) diffeomorphisms.