Hi I’m a mentor for team 373 in the UK.
The UK national event is due on 21 June (our season has not finished yet).
My team have been developing their team props but were unable to access the ML toolchain.
Is the ML toolchain now closed?
Is it possible that it will reopen for UK teams?
If not please can you advise alternatives?
Hello, @richardforster.
We regret to inform you that the Machine Learning Toolchain is down for the season - this closure was announced multiple times in the weekly Team E-mail Blast. Each season FIRST Tech Challenge operates the Machine Learning Toolchain instance from Kickoff to the end of the Championship event, about 8 months. Each month it costs us quite a bit to operate the site (Google Cloud Server Use, Storage, and Data Transfer fees, and of course performing the model training, all which are billed separately), depending on team use. We unfortunately had to make decisions about how to fairly allocate funding for site upkeep and model training use, and only keeping the site operational September through April was the tough decision we had to make in order to give all teams the most utility throughout the competition season.
Normally I would suggest creating your own instance of the tool (using our instructions), but the Google Cloud (which is a required management component) is deprecating the Machine Learning APIs that we utilize and will no longer allow new instances to be created that depend on those APIs. Until we finish the work to move to the new ML APIs (which is much more difficult than we originally anticipated), there is no alternative for teams to use our tools.
On top of that, TensorFlow tools recently changed and the released FIRST Tech Challenge TensorFlow libraries no longer support models created by newer TensorFlow tools (which are what external model building utilities use). We’ve found that merely updating our tools is not sufficient. Until we resolve these issues there is no documented process that I can recommend that will generate TensorFlow models that are guaranteed to work with any released FTC Software Development Kit. This has unfortunately been the case most of this past season, and it was only recently that we were able to discover why. Many of the core technologies we use were deprecated during this past season (that’s the price we pay for being on the cutting edge), and our TensorFlow tools were built around those core technologies.
I’m sorry to be the bearer of bad news, but this is the state that we are in. While the Machine Learning Toolchain is down, we’ve been working to migrate our systems to the new recommended tools. We’ve also been evaluating how we can best serve teams in the 2024-2025 Season, and what that may look like for our Machine Learning and AI tools. We apologize for the inconvenience this causes you, but we hope we will emerge with better solutions come kickoff.
-Danny
Thank-you for the comprehensive and rapid response, I will share with the team.
There is learning for us regarding monitoring Team E-mail Blast.
Our original teamp props are undersized, but I think this might be our best plan B with regard to using a team prop for the autonomous phase of the challenge.
Hello @ddiaz ,
We talked a couple years ago about a tool I created using Google’s CoLab to create TensorFlow models that worked. It was created because my team is registered through a regional partner and so doesn’t have an account with FIRST directly (and so couldn’t sign into the toolchain).
I had updated this tool for the CENTERSTAGE season and it worked perfectly.
I am looking into updating it again for this season as well.
Would you or the FTC team like to take a look and see if this tool could be used for the entire community?
I know that version 10 of the robot controller will not support TensorFlow. Is this because the ftc-ml tool is not useable right now? Or because of other compatibility errors?
Hey Kyle.
We had a number of setbacks force our hand to drop support of TensorFlow within the core FTC SDK. The straw that broke the camel’s back, however, was a breaking change in TensorFlow itself that was not compatible with our SDK support infrastructure. Our current (9.x) infrastructure is not compatible with newer TensorFlow toolchains, and so in order to support teams making models with current tools we would have to make significant changes that we’re not able to make at this time. TensorFlow is a “research project” for Google anyway, I think we were “lucky” to be able to use it for as long as we did.
The TensorFlow Object Detection APIs have significant issues, mostly dealing with the amount of time it takes to train coupled with the significant difficulty to create models that are relatively immune to slight lighting differences at venues. I personally prefer to move to appliance-based machine learning models, for example you can train an object detection model on a HuskyLens in about 20 seconds versus 6 hours with TensorFlow and the ML Toolchain; you can even retrain in the exact lighting conditions in another 20 seconds during calibration/inspection. OpenCV is also a fantastic tool for those who “just want results”, though we do not directly support OpenCV on Blocks (we indirectly support it using java-based myBlocks though).
As we move forward we’ll be constantly evaluating the device landscape looking for better ways to incorporate machine learning, AI, and other similar technologies. But for right now it doesn’t make sense for us to pour resources into technologies that are not directly helping teams (sure, they’re learning a lot about ML and AI, but we need them to learn more than “it’s really difficult to use effectively”).
Love that statement, Danny. I am still unsurprised by Google’s feeble efforts into supporting keyword detection with their voice kit (versions 1 & 2). Even at $5, there were no takers at local retail outlets for V1 when it was superseded by V2.
DFRobot is a nice company. Had the privilege of stopping by their office nearly a decade ago (much to the surprise of my colleagues, who were wondering what was my side gig). Unfortunately, I feel that their solutions suffer from the same mentality. UNIHIKER (the best example) is following the same dead end footsteps of Intel Edison. FireBeetle is nice but not in scope for FTC. HuskyLens would require more space for discussion, although I do see North FiT always exploring FTC solutions.
Thanks as usual for keeping FiT on the vanguard of these solutions.
Regards.