Anki Vector: Autonomous mini robot with charm and its own SDK put to the test

"A toy robot the size of a hamster? What nonsense!"
[Vector rolls into view]
"OMG, he's so cute!!! Take my money!"

This is how many of my colleagues' first encounters with Anki's new mini robot "Vector" went. The chirping caterpillar robot simulates a surprising number of emotions with its lovingly animated eyes, head and shovel arm movements: Vector is happy when it recognizes people, shies away from table edges in fear, coos contentedly when stroked, closes its eyes to sleep and snores peacefully. If it notices movement in its immediate surroundings, it wakes up and moves away from its charging station to investigate the source of the movement or to play with its cube.

In fact, when the mini-robot was released in October 2018, the asking price of over 250 euros was an obstacle for many people to buy Vector. In the meantime, however, the price for Vector has fallen dramatically. The background: manufacturer Anki has been insolvent since April 2019 (see box). Nevertheless, the little robot "lives on".

A long press on its back switches on the robot, which raises its head and opens its eyes after booting. The initial setup requires the Vector Robot app for iOS (or the Vector Robot app for Android). A Bluetooth connection is established with the smartphone or tablet to connect the robot to the WLAN (2.4 GHz), view photos taken by Vector or delete learned faces. Anki justifies the need for a 2.4 GHz WLAN with lower power consumption and greater range than the 5 GHz band.

Once connected to the Internet via WLAN, Vector can manage on its own – as long as it finds its way back to its charging station before the battery runs out. In our test, Vector lasted a maximum of half an hour before it had to make a pit stop. Vector also remains active in the charging station; after ten minutes it is ready for mobile use again.

Technical inner workings & sensors

Vector's predecessor "Cozmo" already had plenty of charm. However, its "artificial intelligence" was entirely contained in its companion app on the smartphone or tablet due to a lack of its own computing power.

Vector's brain, on the other hand, is the Qualcomm APQ8009 processor, whose four Cortex-A7 cores provide enough power for the robot to analyze the data from its sensors autonomously.

Thanks to the built-in inertial sensor (measurement of acceleration, position, rotation rates) and infrared laser scanner, Vector explores its surroundings using Simultaneous Localization and Mapping (SLAM). This works down to the centimetre for simple manoeuvres – as long as the robot is not knocked off course by an obstacle. If an object is placed in Vector's path that it cannot overcome, Vector continues in the wrong direction and does not correct its course accordingly. Apparently, the robot firmware does not check the direction of travel continuously, but only at pre-calculated waypoints where Vector has to change direction.

The infrared laser scanner installed under Vector's 720p head camera for distance measurement has a resolution down to the millimeter range. To prevent Vector from falling from a great height, there are also infrared sensors in all four corners of the underbody. They do a good job; during our test, Vector only once maneuvered itself into a situation where it almost toppled off the desk.

While Cozmo didn't even have its own microphone (the smartphone or tablet had to be used for this), Vector's back features an array of four microphones that use beamforming to determine the direction from which the robot was approached so that it turns to the person speaking.

Also integrated into the back is a gold-colored capacitive touch sensor, which Vector uses to register whether it is being stroked.

"Hey, Vector!" - Voice commands & Alexa

Like Alexa, Siri & Co., Vector needs an internet connection to process voice commands and answer knowledge questions. If you call it with "Hey, Vector!", LEDs on its back light up blue and signal active voice input. Although this takes place in the Amazon cloud, Anki promises – unlike Amazon Alexa – to delete the voice snippets immediately after processing. Anki dedicates a separate FAQ to the topic of privacy and security.

Communication with the Anki servers is always encrypted using Transport Layer Security (TLS), which is why we were unable to look into the data packets. However, according to our observations, Vector only communicates with the Internet when voice input is active. Exception: automatic firmware updates during the nightly maintenance window.

Vector optionally supports the digital assistant Amazon Alexa (currently only with the US English language setting). If you link your Amazon account to the Vector app, you can then access all the digital assistant's functions and skills via Vector – except music and video playback, reading books aloud and telephony.

If you call "Alexa!", Vector's googly eyes are replaced by a glowing blue ring (see video). As long as Alexa is active, Vector behaves like a mobile Echo speaker; its robot functions cannot be controlled via Alexa. As Alexa communicates directly with the Amazon servers, the same data protection concerns apply here as for Echo & Co.

Incidentally, Vector reacts much more quickly to "Alexa!" than to its own activation word "Hey, Vector", but sometimes too hastily. In the test, Alexa responded more frequently than an Echo Dot, for example, even without mentioning the keyword. Conversely, it is more difficult to end Alexa's torrent of words prematurely with Vector; sometimes only pressing the button on Vector's back helps. Unfortunately, the activation word for Alexa cannot be customized on Vector.

Face recognition could be improved

Thanks to its wide-angle 720p camera and built-in face recognition, Vector can in principle remember up to 20 people and address them by name. As the UK import version only "understood" and spoke English (US, EN, AU) at the time of testing, this only works to a limited extent with German first names: "Volker" then becomes "Falcon" or "Voltaire".

Changing hairstyles, glasses, hats or facial hair throw Vector off his stride: a colleague he met with bangs didn't recognize Vector the next day because she had a different hairstyle. Although Vector's camera has a resolution of 1280 × 720 pixels, the image quality is poor, especially in dim lighting conditions, meaning that the robot sometimes fails to (re)recognize you.

Incidentally, the camera stream is only transmitted to the local network, so you don't have to worry about someone watching you from the Internet. The stream can only be accessed via the protected access of the Software Development Kit.

Vectors "Cube" toy cube

Vector's scope of delivery includes a plastic cube marked with various symbols and equipped with four RGB LEDs. Vector can roll it around, retrieve it on command or use it as a starting aid for a wheelie. The symbols on the cube help Vector to recognize the orientation and position of the cube.

Vector communicates with the cube via Bluetooth and can therefore control its LEDs. That's all it can do with the cube at the moment. The cube therefore gives the impression that Anki is only including it so that the shovel arms inherited from Cozmo have a use.

The status of the Cube can be checked via the Vector app. It contains a replaceable "Lady" type battery (AM-5/LR-1), which should last for around 40 hours of play with Vector.

Vector SDK: Programming robots

If Vector were just a cute toy, the price would actually be quite high. However, thanks to its numerous sensors, Vector also offers a good introduction to (robot) programming.

Since December 2018, the alpha version of the Vector Software Development Kit has been available to everyone (registration required). The SDK is based on that of its predecessor Cozmo, meaning that Vector can also be programmed in the popular Python scripting language. The API currently provides 29 Python objects with numerous classes and functions that can be used to control all motors, sensors and the screen.

Installation is very simple: After installing Python (Download), the command

py -3 -m pip install --user anki_vector

to install all the required packages from the Anki repository. Communication with Vector takes place via HTTPS over the WLAN connection; the computer on which the SDK is running must therefore be in the same WLAN as Vector. A certificate from Anki's servers is required for authentication. The command

py -m anki_vector.configure

asks for the access data of the Anki account created when Vector was set up, the name of the robot and its serial number in order to download the certificate.

Anki provides a dozen or so tutorial scripts as well as four more complex applications for control: a 3D viewer, an interactive shell, the proximity mapper and a Vector remote control. All applications and scripts do not run on Vector itself, but transmit the individual commands via the encrypted network connection. The constant need for authentication with the robot makes this a rather sluggish undertaking.

The first steps are very simple; the following script, for example, has Vector follow an outline of a heart:

#!/usr/bin/env python3

"""-------------------------
   Make Vector drive a heart
   -------------------------"""

import anki_vector
from anki_vector.util import degrees, distance_mm, speed_mmps


def main():
    args = anki_vector.util.parse_command_args()

    # The robot drives straight, stops and then turns around
    with anki_vector.Robot(args.serial) as robot:
        robot.behavior.drive_off_charger()

        robot.behavior.turn_in_place(degrees(-30))
        robot.behavior.drive_straight(distance_mm(200), speed_mmps(100))
        for _ in range(6):
            robot.behavior.turn_in_place(degrees(30))
            robot.behavior.drive_straight(distance_mm(30), speed_mmps(100))
        robot.behavior.turn_in_place(degrees(-120))
        robot.behavior.drive_straight(distance_mm(30), speed_mmps(100))
        for _ in range(6):
            robot.behavior.turn_in_place(degrees(30))
            robot.behavior.drive_straight(distance_mm(30), speed_mmps(100))
        robot.behavior.drive_straight(distance_mm(170), speed_mmps(100))

if __name__ == "__main__":
    main()

Vector has been available in the USA since October 2018: the website Kinvert, which specializes in programming courses for children and young people, explores Vector's possibilities and has published various Python scripts for Cozmo and Vector along with explanations, such as a Pong game. The VectorCloud project can also be found on GitHub: a web interface that has been expanded compared to the "Remote Control" example application, including a demo "app store".

In the Anki Developer Forum (registration required) there are advanced projects such as the Vector Explorer Tool –, a port of the Cozmo Explorer Tool –, with which you can easily try out all the hardware and software features of Vector, including game controller support. With the help of the object recognition system You Only Look Once (YOLO), Vector can also play blackjack with real playing cards, as the VectorCards project shows.

Several robots can also be addressed using their serial numbers via the SDK, so that interactions between them could also be realized in principle.

Conclusion

For nerds, Vector is the perfect pet – it only eats electricity, is house-trained, allergy-friendly and (often) even listens to what you say. Thanks to the built-in sensors and the 4-core processor, the little robot is powerful enough to react to its environment depending on the situation. Vector is fun for quite a while. After a few weeks it does wear off a little. However, Vector's life of its own keeps drawing attention to itself, as long as you make sure that it always finds its way back to its charging station.

The Python SDK opens up a whole host of exciting options: It not only lets you add your own functions to Vector, but also learn Python along the way.

In fact, the robot is particularly interesting for many hobbyists right now due to the price drop caused by the Anki insolvency, despite the risk that the cloud services might stop working at some point. It would then no longer be possible to authenticate against the Anki servers in order to run Python scripts on the robot. But this risk seems to have been averted by the takeover by Dream Digital Labs.

Updates

06.05.2019: Warning about Anki's insolvency incorporated.

06.01.2020: Added reference to the takeover by Digital Dream Labs and updated text passages and conclusion with regard to the price drop.

(vza)