Profile cover photo
Profile photo
Tom Hume
885 followers -
Long-time squawk box software chap. Snagged on frowns, and slowly dawns.
Long-time squawk box software chap. Snagged on frowns, and slowly dawns.

885 followers
About
Posts

Post has attachment

Post has shared content
Public

Post has attachment
Just had the joy of seeing this at the SF Jazz Center. Charming live assemblage of a puppet film, cut across 6 different on-stage sets with music from Kid Koala and a struggle quartet. Absolutely beautiful, thoroughly recommended.

+Ellen de Vries this has your paws on it.

Post has attachment
Went to the DIY Robocars event in Oakland for the first time today with my Donkey car. A drubbing for the neural net-based competitors (of whom I was one), who failed to complete a lap even as the computer vision-based teams sped round the track. Much grumbling afterwards, my only consolations being the joy of participation and putting some faces to a few names from the Slack channels. Already looking forward to next month :)

www.diyrobocars.com if you're interested

Post has shared content
Exhibition of robots at the Science Museum. Lots of good stuff - and plenty of creepy stuff - here, with strong representation from Sussex Uni and alumni.

Tickled to see W. Grey Walter's tortoise, unnerved by Telenoid, cheered to see Baxter learning to pick up objects.
PhotoPhotoPhotoPhotoPhoto
7/23/17
23 Photos - View album

Post has attachment
Back in April, I took a half day off work and wandered down to San Jose for the Silicon Valley Robotics Block Party[1]; I’ve been going to some of the excellent SV Robotics events pretty well since we arrived in San Francisco, and this is their annual gala, with contributions from enthusiasts, industry, and everything in between.

On my way out I stopped to watch a talk by Chris Anderson[2], who used to edit Wired and now runs 3D Robotics[3]. He was talking about his experiences building “the largest drone company in the US”, and laid out his theory that most innovation in aerospace over the last 10 years has been at the low end, in consumer drones, because of laxer safety standards (they aren’t carrying people) and lower costs (crashing them is cheaper). He closed by talking about DIY Robocars[4], an effort he’s set up to apply the same thinking to self-driving cars: get hobbyists to build, race and improve smaller, cheaper cars and quickly learn lessons about nimbleness which the larger players in this space cannot afford to learn with their human-carrying, safety-conscious, closely watched efforts.

The DIY Robocars crew get together to race autonomous vehicles each month in a warehouse in Oakland. They have a standard build for their vehicles: a mix of an electric remote controlled car, a Raspberry Pi, and a 3D printed chassis to keep it all together (and a camera in place). Bingo: a self-driving car which is standard enough to let enthusiasts share knowledge and data-sets, and easy enough that a monkey can build it.

I’ve proven that last bit, at least; I finished building mine earlier this week, and took it out to the Panhandle (a park near where we live) to train it yesterday; whilst most folks seem to train them on marked-out racing tracks, I don’t have such a track handy so thought I’d try it on the freshly tarmacced paths which run across a section of the park. I guided the car around a loop of the park by hand, about 8 times, with it recording the journey, then went home and had my laptop chug away analysing the footage and building a model the car could use to drive itself.

Turns out it works pretty well. Here’s[5] a video of it running all the way round the loop (four corners, two long sections and two short, with a mix of sunlight/shade, traffic going along at each end and the occasional pedestrian or cyclist) by itself.

If you’re interested in this stuff, I can thoroughly recommend checking out the DIY Robocars site and giving it a go. I’m a long way from being comfortable working with hardware myself (I’ve not soldered in 20 years or so), and even I found it straightforward: the software is simple to get going; the documentation is mostly decent, and the Slack channel is full of exceptionally helpful people.

Next steps: clean up my training data, gather a ton more, and see if I can get it working more reliably and faster around this track; then see how this generalises to the rest of the Panhandle. In particular I’m wondering how much of the navigation is fixating, beetle-style[6], on the buildings in the distance rather than the road immediately in front of the car. Also if there’s a style of training to be had where I can run the car basically autonomously, but then just correct it when I start to see it making obvious errors…

[1] https://svrobo.org/robot-block-party/
[2] https://en.wikipedia.org/wiki/Chris_Anderson_(writer)
[3] https://3dr.com
[4] https://diyrobocars.com
[5] https://goo.gl/photos/jAgWMHgNhEYHpSTL6
[6] http://www.npr.org/sections/thetwo-way/2016/05/13/477963290/dung-beetles-navigate-poop-pile-getaways-using-celestial-snapshots

Post has shared content

Post has attachment
"Federated Learning enables mobile phones to collaboratively learn a shared prediction model whilst keeping all the training data on device, decoupling the ability to do machine learning from the need to store the data in the cloud".

If you're interested in privacy-responsible mechanisms for machine learning, I thoroughly recommend you read both this blog post, and some of the papers linked from it.

Post has attachment

Post has attachment
We popped into the Smart Home exhibition at Target today. It's a display of technology from a ton of different partners: from the obvious SmartThings from Samsung, through door locks from August to some fairly ridiculous items: a smart basketball, networked ear thermometer, and the Parrot plant monitor (which I'd probably use).

Kudos to them for not just highlighting the big brands, but the resulting picture of IoT, and the way it was presented, was a dreadful cacophony of needy devices, screaming for attention, shouting out their cleverness with twee "personalities", and relying on an bewildering set of apps and infrastructure to configure and use. The silver lining: after seeing all this, efforts like Thington seem less like nice-to-haves and more essential.

www.thington.com
PhotoPhotoPhotoVideo
Photo
5/22/16
5 Photos - View album
Wait while more posts are being loaded