I’ve been asked multiple times now how to sync animations and speech on a NAO – or Pepper for that matter; especially from Python.
The answer to that is, there are two options:
The first one is to create the animation in Choreograph and then export it to a python script. You then create your usual handle to the text-to-speech module, but instead of calling the say method directly, e.g., `tts.say(“Hello”)`, you call it through the module’s `post` method, e.g., tts.post.say(“Hello”). This method exists for every function in the API and essentially just makes a non-blocking call. You can then call your animation.
You create a custom animation in Choreograph, upload it to the robot, and call it through AnimatedSay or QiChat. Other than being the, I think, cleaner solution, it allows you more fine grained control over when in the sentence the animation starts and when it should stop. This is what I will describe in more detail below.
Step 1: Create the Animation
Fairly straight forward, and the same for both solutions. You use Choreograph to create a new Timeline box in which you create the animation that you would like. You then connect the timeline box to the input and output of the behavior and make sure it works as you’d expect when you press the green play button.
Step 2: Configure the Project and Upload it to the Robot
In this step, you configure the new animation to be deployed as an app on the robot.
Go to the properties of the project.
Then make sure to select a minimum naoqi version (for NAO 2.1, for Pepper 2.5), the supported models (usually any model of either NAO or Pepper respectively) and set the ID of the Application. We will use this when calling the animations, so choose something snappy, yet memorable. Finally, it is always nice to add a small Description.
Next, we need to reorganize the app a bit. Create a new folder and name it after your animation; again, we will use this name to call our behavior, so make sure it’s descriptive. Then move the behavior that contains your animation – by default called behavior1.xar – into the folder you just created, and rename it to behavior.xar .
Finally, connect to your robot and use the first button in the bottom right corner to upload the app you just created to your robot.
Step 3: Use ALAnimatedSpeech from Python
Note:If you don’t want NAO to use the random gestures it typically uses when speaking in animated speech, consider setting the BodyLanguageMode to disabled. You can still play animations, but it won’t automatically start any.
For existing animations – that come with the robot by default – you call the animation like this
"Hello! ^start(animations/Stand/Gestures/Hey_1) Nice to meet you!"
Now, animations is nothing but an app that is installed on the robot. You can even see listed it in the bottom right corner of Choreograph. Inside the app, there are folders for the different stable poses of NAO like Stand, or Sit, which are again divided into types of animations, e.g., Gestures which you can see above. Inside these folders there is, yet another, folder named after the animation (Hey_1), inside of which is a behavior file called behavior.xar.
We have essentially recreated this structure in our own app and installed it right next to the animations app. So, we can call our own animations using the exact same logic:
"Hello! ^start(pacakge_name/animation_name) Nice to meet you!"
It also works with all the other aspects of the ALAnimatedSpeech module, so ^stop, ^wait, ^run, will work just as fine. You can also assign tags to your animations and then make it choose random animations for that tag group.
Our lab owns robots build by SoftBank that we use for experiments; we have a Pepper and some NAOs. At the moment, I’m working on a NAO.
They are quite pretty robots. I mean, they can barely walk around, let alone navigate the environment, they can’t do proper grasping, the build-in CPU is so slow and hogged by the default modules, and streaming video from the robot for remote processing happens at about 5 FPS. So you can’t really do any of the things you would expect you can do, but hey, they look really cool 😀
Okay, jokes aside, the manufacturing of the robots is actually pretty solid. Being able to get your hands on a biped for about 6000€ is solid, and, despite some stability issues, it can walk – however, nobody really uses that feature in social robotics research. They also come with a huge sensor array, that makes every smartphone jealous. Hardware wise both, NAO and Pepper, are good robots.
The thing that is lacking – by a landslide – is the software. The robots come with an API, but that API is proprietary – in itself, not a problem. The problem starts where the documentation ends. Documentation is shaky, disorganized, not very clear, and – for all the cool parts – nonexistent. In short, you don’t get to read the code and you don’t get good documentation to help you either; hence, if something breaks, you are blind and deaf somewhere in the forest of code and have to find the way out yourself.
Pepper can grasp, it can do navigation, and you can stream video data at a decent FPS – the same is true for NAO; it can do all the things I just complained about. You just have to write the code yourself.
This is what I will talk about in this post. I will not go into grasping or walking, but we will look into navigating the joint space more efficiently. That is, we will have a more in-depth look at ALRobotPosture, some of the hidden / undocumented functions, and how we can use this module for some pretty sick motion planning.
Note: Everything in this post works for both NAO and Pepper. For ease of reading, I will only reference the NAO, because – I think – that is the robot most people reading this will own.
ALRobotPosture, an Overview
If you own either a NAO or a Pepper, you have probably noticed that, when you turn (and autonomous life activates) it on, it moves into a certain pose. For Pepper, it is always the same, for the NAO, it depends if it was turned on sitting or standing. This is RobotPosture in action. The same is true after we play an animation. Once it finishes, NAO moves back into a specific pose, waiting for the next command.
This is the most visible action of the module. When no other movement task is running, it will move the NAO into a stable position. The other thing it does, is it transitions between these stable poses. For example, when you want NAO to either sit down or stand up, then it doesn’t play an animation. It actually uses RobotPosture to navigate the joint space from one stable posture to another until it reaches the Sit or Stand posture respectively.
In essence, RobotPosture is a list of configurations – points in joint space – that serve as stable positions the robot can move into. These points are connected; there is a neighbor relationship between them. They are also attractive; hence, when no other motion is running, NAO will move into the closest posture (closeness being defined as closeness in joint space).
The interesting part is that movement between poses is not done as a direct line in joint space. This could be rather dangerous, since the robot would just fall, if it would move in a straight line from the Stand to Sit. Instead, planning is done in the topological map – a directed graph -, that is defined by the poses and their neighbors. NAO then moves in a (joint space) direct line from the current pose to a neighboring pose and goes through different poses until it reaches the final, desired pose.
I visualized the standard poses in Figure 1. Additionally, there is the USRLookAtTower pose, which is a custom posture I’ve added for a project I’m working on. You can also see it in the picture I chose for the beginning of the post. It looks a lot like normal sitting, but the head is tilted downwards. I will walk you through how I did that in the next section. I also color coded the sitting and standing postures, because they are the most used – but mainly because it looks nice 🙂 .
The graph is laid out using force-directed graph drawing, where the force corresponds to the euclidean distance between nodes in joint space. However, I took a bit of liberty to prevent label overlap. As you can see, there is no direct connection between Sit and Stand; the robot would move through unstable territory. (We could, however, add such trajectories ourselves, creating a fast, dynamic stand up motion – e.g., for robot football.)
Another advantage of this approach is that it is very computationally efficient. Since we have an abstract map of how poses are connected, we can quickly figure out if a pose is reachable, and compute a path to that given pose.
Enough theory, show some code already! Okay … okay. Here is how to use the basics of the module:
The snippet will make the robot run through all the available poses and announce the pose’s name once there. This is about the best you can do with the official part of ALRobotPosture; not that much.
There is a lot more functionality in the module. There just isn’t any documentation of it on the web. We can look at all the methods in a module via:
Alternatively we can use qicli (with the parameter –hidden) to list all the functions in a similar fashion. Qicli is documented here.
Here we can find a few very promising functions:
_isRobotInPosture(string, float, float)
This function is similar to getRobotPosture(). However, instead of giving the current posture, it gives a boolean that is true if the robot is in the given posture. The two floats are threshold values for the joint angles and stiffness. That is, by how much is the current pose allowed to deviate from the defined pose for us to consider them the same.
It returns a triple of (bool, [bool] * 26, [bool] * 2) on a NAO robot. The first boolean tells us if the pose has been reached overall, the second is a breakdown if the pose has been reached for each joint. Finally, the last array is the same for stiffness.
This function is useful if two poses are close together. In this case getRobotPosture() may not show the correct pose; however, we can still differentiate with _isRobotInPosture().
Make your own network of postures, export it, use this to upload it to an army of NAOs, and dominate the world.
Given a serialized graph of poses, it will load it and replace the current posture graph. The string is the (relative) path to the file. It returns a boolean indicating if the loading has succeeded.
Important: The file path is relative to ~/.local/share/naoqi/robot_posture on the robot, so the posture file has to be stored in that directory on the robot.
This is a strange one. While not immediately useful to us, it will re-generate a Cartesian map that the module uses internally to navigate between poses. You have to call this after loading a new posture library or adding individual postures. Otherwise the new postures won’t work!
Pretty self explanatory. Look up the id of the posture using it’s name. Takes the name of the posture and returns an integer that is the id.
Takes a posture id and returns a boolean. True if the posture is reachable from the current robot pose.
The string is the name that we want to save the file as and it will be saved under ~/.local/share/naoqi/robot_posture on the robot.
_addNeighbourToPosture(int, int, float)
Adds a vertex to the graph pointing from the first posture (indexed by the first int) to the second posture. The third value is the cost of traversing along this edge, which can be used for more sophisticated path planning.
Saves the current pose as an edge with ID int and name string.
Putting all these together, we can create custom poses as follows:
Use the Animation Mode (or any other method) to move NAO into the desired pose
_saveCurrentPostureWithName() to add the node to the graph
_addNeighbourToPosture() to connect it to the graph (edges are directed! we have to add both ways)
export the postures via _savePostureLibrary() (this will save the file in the correct place)
In our code: import our custom poses using _loadPostureLibraryFromName()
re-generate the cartesian map _generateCartesianMap()
Here is a code snippet that adds a custom posture called “myPosture”, exports the library, imports it, makes the robot sit down, and then go into “myPosture”.
And just for good measure, a video showing what the robot does when running the snippet:
Naturally, you can be more fancy with this. I am particularly excited about the possibility to do dynamic movements, i.e., one-way trajectories. However, my supervisor will probably kill me if I actually dabble in this area, because the chances of breaking a NAO like this are … elevated.
I hope this article is useful. If you liked it, feel free to leave a like, comment, or follow this blog! I will keep posting tutorials in the area of robotics, AI, and social robotics research.
A few weeks ago, I ran my first pilot study on Amazon Mechanical Turk (MTurk).
Essentially, MTurk is a platform that you can use to label data and get participants for studies (using money). There is of course more that can be done here, but from my current understanding, these seem to be the two main uses in academia.
The completion of a survey usually takes the following form:
A worker / participant on MTurk accepts the work and is presented a link to the survey.
The worker copies his unique, anonymized workerID into the survey (so that you can reference the data)
The worker completes the survey and is presented a unique random code at the end
The worker copies the code into a form on MTurk’s website.
You match the IDs of workers that completed the survey with IDs in your database and reward workers for their time
You used MTurk before? This sounds familiar? Yes – however, two parts in this chain are rather weak: (1) the manual copying of the workerID and (2) the generation, matching, and copying of the random code at the end. If either of these fails, you have to work out manually if a worker has participated or not.
As you have already guessed, there is a better way to do this (why else would I be writing about this? :D). MTurk offers to host a so called ExternalQuestion, which allows you to embed a custom website via an <iframe>. On top, it passes some meta information, such as the workerID, to the website; all you need to do is read that info and use the value as you see fit. Additionally, it allows the website to submit a form back to MTurk as proof of having completed the work, which we will do at the end of the survey.
In short, we integrate the survey directly into MTurk thereby getting rid of above pitfalls. It roughly follows these steps:
Create the Survey on the survey tool (in this case Questback)
Create a small adapter website on GitHub (may not be needed if you have a different survey tool)
this will accept and forward requests from Mturk
allow you to create an arbitrary preview of the survey
pipe the form back to MTurk when the survey is finished
Add URL parameters to the survey
forward people to the adapter website upon completion
Host an external question pointing to the adapter website on GitHub
Small Adapter Website on GitHub
At first, I tried to link Questback and MTurk directly; then I discovered two limitations making this impossible: (1) Questback only accepts URL parameters named “a=..&b=..&c=..” instead of full variable names, and (2) Questback can not post the results of a form; I could only find forwarding via GET.
Hence, I set up a small website hosted on GitHub to do the plumbing between both websites.
Note: If you are working with the sandbox, you have to change the mturk_url appropriately.
This does 3 things, depending on how it is called:
if the URL contains a parameter called “returnToMTurk” then we assume that the Questback is forwarding to this website via GET. In this case we take the payload (in this case attentionCheckPassed) and forward it to MTurk as a POST request.
if the URL contains ASSIGNMENT_ID_NOT_AVAILABLE it means that the survey is being previewed. In this case we do nothing and show this website, which will act as the preview of the survey (e.g. display a screenshot of the hit or some other, relevant information to inform people what this task is about). Note that you want to avoid people submitting your survey at this stage, so showing them the raw survey may be counterproductive at this stage.
Otherwise the website is called from MTurk by a worker who wants to complete the survey; in this case we rename the URL parameters from their actual names into a, b, c, d and forward the request to the Questback survey.
Add URL parameters to the survey
In Questback this can be done in the survey properties > User-defined variables.
As mentioned before, the URL parameters will have names a,b,c, … and can be accessed in the survey tool via #p_0001#, #p_0002#, #p_0003#, … .
Forward People to the GitHub Adapter Upon Completion
This is easily done in the properties section of the final page under Questionnaire editor > Final Page > Properties > Redirect to Survey .
The address should point to the GitHub Adapter and the URL should include three parameters: (1) the assignmentID sent from MTurk, (2)-this is reallyimportant– at least one additional parameter to store in MTurk as the result of the task and (3) the “returnToMTurk” parameter used to tell the adapter what to do. An example of a URL could look like this
Note that the assignmentID is set to #p_0001# which is the the first parameter (“a“) passed to the survey from MTurk. In above example attentionCheckPassed is a variable from the survey which we used to determine if participants payed attention or just mindlessly filled out the survey. This is an aggregate of multiple questions and computed by the survey tool upon completion of the survey. This can later be used to automatically accept / reject / ban workers that have completed the assignment.
It is also important to note that MTurk expects the assignmentID and at least one additional parameter to be send through the form’s POST request. For some reason, the additional parameter is mentioned nowhere in the documentation, but, instead, tacitly assumed.
Additionally, the checkbox next to the phrase Automatically add ospe.php3 to URL and Add return ticket have to be disabled. You would use these if you were forwarding / returning to another ESF survey; this isn’t the case here.
Host an External Question Pointing to the GitHub Adapter
All that is left is to actually host a task on MTurk. In this case an External Question.
Unfortunately, this is currently impossible to do through the web UI. Hence, we have to use the API; I decided to do it in Python by adapting a code snippet that I found on the web. It consists of two files: (1) config.py, which stores the credentials, and (2) create_hit.py which creates the actual hit.
That’s it. Running the code will create a HIT that will point to the website on GitHub, which itself points to our survey. The survey will point back to the GitHub website, which will point back to MTurk, going full circle. Neat!
As always, I hope this is useful to some of you and feel free to drop a comment or reach out to me if you have questions.
For the love of all things good; there is two things I don’t like: (1) unnecessary convoluted setups and (2) redoing work I’ve done earlier. SSL Certificates seem to combine both into one beautifully painful mess.
I once again found myself in the need to generate a throwaway SSL certificate for local development and testing. As I’ve posted earlier there is a way to do that with more or less effort. Since I found myself doing this for the third time now I decided to spice things up a bit, cutting down the time it takes. I’ve created a small docker container that’s sole purpose it is to create a certificate from a given config file. The way it works is:
Create a folder (e.g. ./certs) and place the below config file (named config.cfg) into it.
Modify the config file, adding all the SANs required for this certificate
docker run –rm -it -v./certs:/certs firefoxmetzger/create_ssl
That’s it. It will drop a `private.key` and `certificate.crt` into the previously created folder. It will also print the properties of the certificate into console so you can make sure the SANs are actually added.
Here is the config.cfg template
And here a link to the GitHub repo with the code of the docker image:
The second problem in this years Code Jam was Trouble Sort. This problem was substantially easier then the first one. At least for the small test.
The idea is that one performs bubble sort with three elements at a time and, if the last element in the triplet bigger then the first, one reverses the entire 3-element sub-list. Pseudo-code for this algorithm was given. The goal of the exercise was to, given a list, sort that list using above strategy and then asserting that it was ordered. If it was one should print “OK” otherwise one should print the index of the (0-indexed) first element of non-increasing order (e.g. [3,9,7] should return 1).
The problem with passing the big test was that lists could have up to `10^8` many elements and above trouble sort implementation has `O(n^2)`. This means using the naive implementation will quickly run into the timeout.
The “trick” is to see that trouble sort is equivalent to splitting the unsorted list into elements with even and odd index, sorting them and then concatenating those sorted lists. Since this sort can use a reasonably fast algorithm (e.g. pythons native one) it can happen in O(n*log(n)) which will be significantly faster and work on the second, big set of tests. (I know this because I read it in the analysis).
The way I got that intuition was thinking about the fact that reversing a three element list is the same as swapping elements i-1 and i+1 . So we swap with a “distance” of two which made me wonder if it would be possible to swap the second and third element in the list in some way. The only swaps that involve the third element are when the “cursor” is at position two and four and that would swap with the first or fifth element respectively which are both uneven numbers. This trend continues for all numbers (even swap with even, odd with odd) and suddenly you begin to see a pattern.
Unfortunately I failed this challenge completely. My logic is correct, however if you look carefully at the last line, you will see that it outputs something like “Case 1:” however it should output “Case #1:”. I spent over 2 hours trying to find the mistake…
Sometimes it’s the small things that catch you. This probably won’t happen to me again … ever.
During last week I’ve learned about a Code Jam hub in our University. The idea is to meet up and participate in the Google Code Jam 2018 . I ended up participating and advancing into the first round, so I thought I’d share my solutions here for anybody interested. (And for me for further reference)
An alien robot is shooting a beam that will destroy all algorithms knowledge (not sure how an algorithm can have knowledge, but lets not get philosophical).
The beam starts with strength of 1 and follows a given sequence of actions (string with two literals “S” and “C”). Whenever it shoots “S” it deals damage equal to it’s current strength and whenever it charges “C” the strength is doubled. The sequence “SCSS” would thus deal 5 damage.
You have a shield that can absorb D damage (humanity needs the D!) and can swap any two adjacent literals in the string. What is the minimal number of swaps to reduce the damage dealt to a number <= D? (If impossible return “IMPOSSIBLE”)
You can understand this problem as a tree search (again! gives you a competitive advantage to know these things). The state is the current string and it’s children are all the new states that result from swapping in different places. For each such state you can calculate the damage. To find the closest node to the root node that has damage <= D a possibility is to use simple breadth first search.
Breaching from my usual style I will post the code first and then talk about it below. This is the pure, unmodified code I submitted during the hub, so it is intentionally left a bit messy. It’s still fairly readable though … It’s python and it’s short.
Note: This will work for “small” sequences and is too inefficient for large ones. The branching factor of the tree is len(sequence)-1 which can be up to 30. I timed out on the second (big) test, because I didn’t optimize this part.
Pure breadth first search on a tree with branching factor >10 is generally not a good idea. As such I’ve implemented the search using a queue rather then choosing a recursive approach. This allows to better scale the code as ideas for run time optimization come in.
By optimization I mean pruning of nodes that are clearly disadvantageous. One example is in line 32 where I keep a dictionary of nodes that have been visited. If I encounter the same state again, I already know a shorter path to it; thus I can safely ignore it. To be honest, I should have done the weeding out directly in get_nodes to reduce constant overhead.
Another optimization, which I didn’t implement, would be to see that this is a convex problem. There is exactly one clear minimum (S[…]SC[…]C) and one clear maximum (C[…]CS[…]S) along the border of the problem and there are no local optima in between. Thus, we can ignore all swaps that increase the damage or keep it constant, i.e. “SS”, “SC”, “CC”, and focus on those that swap “CS” into “SC”. This will severely decrease the branching factor. (Note: We already ignore “SS” and “CC” as part of above optimization)
Another consequence of this is that we can swap the queue type from a FIFO (i.e. breadth first) into a priority queue, sorting by current damage dealt choosing the node with lowest damage as the next node. This is essentially Dijkstra’s algorithm (as we don’t use a heuristic, otherwise it would be A*). It keeps amazing me again and again how easily you can swap between different tree searches by simply changing the queue type.
Applying those three optimizations should cut down search time by enough to cope with the big test. It also gets very close to the other solutions I’ve seen; that is “pass over the string back to front and switch all “CS” into “SC”. Repeat until no change was made for an entire sweep of the string and then output the total number of swaps. My approach is simply more formal and can be applied to other problems more easily.
I’ve been looking at the AlphaGo:Zero network architecture  and was searching for existing implementations. I’ve found quite a few (here , here and here) with varying degrees of completeness. The cleanest is probably this one but it depends on Jupyter.
What surprised me was that I couldn’t find one that used Keras’ sequential API. While residual blocks aren’t exactly sequential, from a high level view the architecture itself is; it simply stacks (a lot of) residual blocks. So it should be possible to create something like this, right?
The answer is, of course: Yes, there isn’t much that you can’t do in Python. We are actually using this strategy already. Sequential itself inherits from Layer and, in fact, Container (a class sitting between Sequential and Layer in the inheritance hierarchy) states so itself: A Container is a directed acyclic graph of layers. It is the topological form of a “model”. A Model is simply a Container with added training routines. (source)
It works by defining the residual block as a new Keras layer. Depending on how tightly integrated you want it this can be quite short:
Inside the block we fall back to the functional way of stacking layers. If you want better integration, e.g. model.summary() showing the number of trainable weights, there is additional plumbing. Above just shows the gist . . . (gosh! That pun was bad).
Once that is written, we can use model.add( Residual(32, (3,3) )) as we would any other layer. Nice!
To close with an example, I modified the Keras CNN example on CIFAR10 and replaced the hidden convolutional layers with residual ones. I haven’t optimized performance, but you can see how it works. If you are familiar with the example, you might appreciate how similar it looks.
 Silver, David, et al. “Mastering the game of go without human knowledge.” Nature 550.7676 (2017): 354.