Clicker Training Your Dog Across the Internet: Rover@Home

13 Feb 2006Steve Schwarz

The other night there was a replaying of the PBS show Scientific American Frontiers episode Pet Tech. It was an interesting program and there were two dog related segments that really caught my interest (click here to watch them both online - select the Virtual Dog Training video) . The first segment describes a system to train trainers to clicker train a virtual dog, this would be great to allow working on clicker technique before practicing with your own dog. This article describes the second segment which highlighted the Rover@Home system for remote Clicker Training and my thoughts on how it is even closer to reality since this research was performed.

Remote Clicker Training

Benjamin Resner, a Masters candidate at the MIT Media Lab, came to the realization that high speed internet connections and internet video conferencing software/hardware has advanced enough that you can put together a relatively low cost system for remotely clicker training your dog. He named the system Rover@Home.

A very detailed account of this project is presented in Resner’s Master’s thesis “Rover@Home: Computer Mediated Remote Interaction Between Humans and Dogs”. As you would expect from a thesis, it has a lot of background material, discussion of the asymmetry of the computer interfaces needed by humans and dogs, discussion of the challenges of remote training and a thorough description of his solution. This is the only “book” I’ve read that equally discusses theory and application of both technology and dog training; both aspects of AgilityNerd. I found his thesis interesting and a pretty good read.

Rover@Home Overview

Rover@Home runs in a browser to allow the human to interact with their dog from any computer anywhere in the world. Resner developed a system where an internet connected home computer could be used to capture video of the dog in the training area via webcam, the pc’s speakers could relay the trainer’s commands to the dog, and the pc’s microphone could pickup the sound in the training area. That is all basic video conferencing functionality provided by software like Microsoft’s Net Meeting. With video conferencing you could be at work and monitor your dog via video and sound and talk to your dog over the speakers.

But video conferencing isn’t enough to remotely train your dog. You need a way to reward your dog when they successfully perform a behavior. Resner built a kibble dispenser that could be connected to the home pc and triggered over the internet from the remote computer. The remote computer runs a web browser that connects to the home pc. The page sent to the remote pc embeds NetMeeting and has icons for sending a click sound to the home pc and dispensing a treat. They also modified two toys that the dog could interact with that could trigger responses both locally and remotely. They had an Alley Oop and a squeaky plush panda equipped with sensors that would let the remote trainer know about the dog’s interaction with the toys.

Thesis Review

Resner’s thesis contains a lot of information. I read the PDF on the train and took the following notes that might be of interest.

He has cast his thesis in the realm of Human Computer Interaction. With a human interacting with the remote computer and a dog interacting with the home computer. A key to his analysis is the asymmetry of the interfaces - dogs are not little furry people. Consequently, the interface to the dog is not the same as the interface to the human. The goal is not to make a videophone for dogs.

"We introduce the concept of asymmetric interfaces that address the unique sensory, cognitive, and motor skills of each creature. In designing an interaction between two different species, we emphasize the importance of catering to both creatures equally and not making one creature adapt to the other, or adopting a "least-common-denominator approach."

He goes on to analyze the capabilities of humans and dogs for interacting with their environment and how those senses might apply to interacting with a computer. For example, speakers make useful output devices to dogs, they can hear in the same frequency ranges as humans. However, computer monitors are not as useful due to the flicker of the screen, and dog’s poor perception of images over video. Similarly, smell and touch are not easily provided by technology.

"Because dogs and humans have very different input and output systems, interfaces for each creature will be correspondingly different."

So Resner determines, as you might expect, that you can’t build a fully generalized internet interaction between dogs and humans. So instead he decided to focus on a ritualized task that could be addressed by available technology. This was a key step in this work. Choosing clicker training as that ritualized interaction was a great idea.

Section 3.8 has a good quick description of clicker training. Clicker training is highly ritualized (a key to its success in this project) and the interaction between dog and human can be transmitted over the internet. No contact between human and dog is required by this ritual.

I don’t believe it was discussed on the TV show but one of the challenges of remote training is the poor resolution and time lag of the video getting to the remote location. Furthermore, successful Clicker Training relies on the trainer being able to mark (click) the desired behavior at the precise time it occurs. The time delay of the video being seen by the remote trainer and the time for the “click” noise (say made by a mouse click on the remote pc) to be transmitted back to the dog could be a couple seconds, far too late to be useful.

To address this Resner used the home pc to accurately detect the dog’s interaction with instrumented toys and issue clicks and treats (all wiring to the toys was very low voltage to avoid the possibility of injuring the dogs). The Alley Oop was instrumented with a tilt sensor and a stuffed panda was instrumented with a squeeze sensor. The remote interface allows for “arming/triggering” the toy, allowing for auto click and/or auto treating when the dog interacts with the toy. Once triggered the toy is inactive until rearmed by the trainer.

Instrumented toys can also reduce the need for sending high quality video to the remote pc. By sending the very small amount of data to indicate that a toy was tipped or squeezed back to the remote pc, the video becomes less important. Although if you were training a dog to touch the Alley Oop with it’s nose and not its paw you couldn’t accurately determine that without video information. Resner points out these lower bandwidth interactions could allow the remote pc to be a handheld computer or a cell phone.

Luring to the instrumented toys is supported by incorporating electronic lights and beepers in the toys. However, luring using these toys requires low latency and sufficient video resolution for the remote user to detect the dog’s movement/gaze and reward accordingly. While arbitrary luring isn’t possible, luring to specific objects is possible. Also since there is daily contact with the dog the training environment can be modified to allow training different tasks with different objects.

Resner conducted field trials at a local dog training center and had a number of insights:

  • One trainer started by "charging the speaker" from which the click sound is produced. This helped the dog understand the click sound from the speaker was really the same click as from a local clicker.
  • Some dogs were getting cues from the whirring sound of the dispenser preceding the treat, this could be as good a cue as the click itself.
  • I was surprised to read that some trainers didn't generalize the computer interface to their normal clicker training model. They took some time to understand the click and treat through the computer had the same effect as clicking and treating directly with the dog. This shows my own computer interface bias.
  • Single speaker giving a point source of sound was less confusing than a diffuse stereo sound.
  • Some dogs oriented to the feeder as they would to the trainer who dispensed the treats.
  • There is the potential to extend the system to use multiple speaker output to give the dog spatial cues like "over here" with the sound coming from a specific speaker.

So transference of the human-dog training interaction to the remote trainer-machine machine-dog interaction required a transitional period for both the trainer and the dog. The environment is artificial to the dog and for success Resner had some recommendations for bridging from the handler’s physical presence to their remote presence. These included putting the speaker at the feeder (which was also a good location for the camera for the human), “charging” the feeder, “charging” the speaker, and performing colocated training with the home pc clicking and dispensing the treats.

I’d need to check the thesis again but I don’t believe he actually tested Rover@Home over an actual internet connection. But I don’t doubt the issues he raised would occur.

As described in his thesis I feel the system was very close to being completely successful. I believe additional work could have demonstrated Rover@Home as a viable training environment for a clicker savvy dog and trainer.

DIY Rover@Home

A lot has happened in the software and hardware realm since 2001 when Resner did this work. Computer hardware is much faster, high speed high bandwidth internet connections are ubiquitous in the US and at reasonable cost, high fidelity computer speakers are inexpensive, and webcams with higher resolution and features including tilt-pan-zoom are also available.

Another factor that tips the balance is Sharper Image is marketing a remote controlled treat dispenser called Treat & Train Professional Dog Training System. It comes with a remote for triggering the dispenser and the dispenser makes a beep sound (think click) when activated. I picked one up just for the fun of it on sale this past holiday season for $39 US. The remote activation also removes one more set of wires from the training area. Of course you’d have to hack the remote control to connect it to the PC.

Resner’s thesis includes schematics and parts lists for the hardware he developed. This would is very helpful especially for interfacing the Alley Oop and any external toys. Although there are a number of off the shelf RS-232/USB digital I/O prototype boards available (i.e. QualityKits) for the budding hardware developer.

On the software side a webserver or app server could be run on the home PC (just like Resner did) and there are a number of options depending on your programming knowledge (e.g. Apache HTTP or JBoss). Once setup with authentication, you need to write the code to interact with the digital I/O and send HTML to the remote PC. Non-trivial code but the fun kind of stuff to develop. Besides writing code that makes hardware do things is always fun (I used to write “soft” real time factory monitoring and control software and before that programmed robots).

I don’t know MIT’s patent claims on this research, but I imagine if one was to build such a system for personal use they might not care (insert usual legal disclaimers here).

Other Directions

The great thing about “rolling your own” remote training platform is the additional directions in which you could take it. One extension could be the computer recording of the training sessions for later review. The remote webcam software usually has this as an option. Annotating each training session as part of the interface (or recording the trainer’s audio with the option of not sending all the audio to the dog) is another direction this could take.

If I had this kind of a system at home I could ask some of the better trainers out there to remote train our dogs from the comfort of their homes. Not only would I have dogs with new fun behaviors, but they’d be intellectually stimulated throughout the day. Well just have to put this on the list of things to do…

Related Articles: