To be fair – I don’t think he’s upset at the devs directly. Clearly Unity is having a hard time figuring out what to do internally, and it is frustrating as a developer when a company you may be trusting your livelihood on begins to hem and haw on what it will or won’t support, and can’t even give you a clear idea of what it’s current thinking is. This is general anxiety – not a direct attack on the Unity devs.
This anxiety is part of the “dust settling” process. Unity’s direction (and silence) could affect his livelihood (and will affect the livelihood of many others). Maybe someone important at Unity will hear these fears and frustrations and try to ease them by offering people something more concrete. Or maybe Unity will just implode first because nobody there listens.
This is a bit of a reach. The team lead who promised to let us know something might have been canned or just relocated to some other (unrelated) team. Again, it’s part of the “dust-settling” process. Expressing one’s feelings and opinions is very much a necessary ‘evil’ to ensure those reading know that there is something wrong. Expressing passion (through fair arguments) matters in the enlightening process. It is on Unity to address them (or get what’s coming for ignoring the dangers inherent in not addressing the reasonable arguments). Ignoring people tends to look waaay worse for a company than simply saying “I don’t know – but I will try to find out”. But that part is just my opinion. Silence, however, isn’t a great option when you don’t know what to do. People expect to be able to trust companies to be genuine and upfront. They very often aren’t – but that’s a deeper problem with Capitalism, and out of the scope of my reply.
That being said –
This part is definitely true. At the very least, one can work with what they have. But in a field where things are moving at a breakneck speed, it is reasonable to expect Unity to pitch-in and work with advancing this field too. But again - that is just my own opinion.
To give credit where credit is due - the old ML-Agents (and Barracuda) team created a really incredible product. This is speaking as a developer who’s fully productionized ML-Agents in a game. Granted it’s rough around the edges in some regards and needed lots of work for efficiency/overhead, no one else has provided a product like this, that can be easily deployed in a game, yet. If you spend anytime in the repo (C# and Python) you’d see that it wasn’t an easy feat to achieve.
It’s sad to know that a lot of talented people were let go, but as is, the product is still intact and is in a good place for others to use and for a new team to jump in and keep working/developing in.
I wish the old team the best and hope that we can turn the outlook of ML-Agents into something hopeful and less negative. Even with the old team gone, we are a community that can continue to innovate, share insights and developments, and provide useful feedback for future tech/improvements.
This is true. Looking into this code myself and, personally, I’m very grateful for their efforts!
I agree with you on this for the most part.
I feel like if we make any improvements on the “rough around the edges” parts, we should all give back to the memory of the team that made this possible by sharing our changes / upgrades to the process on github, provide documentation, or just give back in some way to make others’ lives a little easier, so that we won’t all have to keep reinventing the wheel so much. But that’s just me…
http://www.siggraph2022.unity.com/
They have one session about AI and Physics Assisted Character Pose Authoring on Wednesday 5:10pm which is using ML agents…
The GitHub repo has not been updated with the latest changes. The last commit on the public repo was in January, but there was a new version released on PyPI in April that includes code changes that have not been put onto Git. Not sure if this was an oversight or what, but since it is still under the Apache License, I have copied the updates to my own fork so we can continue to maintain it if the Unity devs don’t want to.
I don’t think so. And I also don’t think there ever will be either. Not unless Unity changes direction as a company - and fast.
Unity, as a company, seems to be in its death throes right now. When your whole Visual Scripting (and most other) projects go dark, instead of developing your core product further, you decide to diversify your income, and then, when that fails, you decide to sell your soul (i.e. a merger) to a subpar advertising agency responsible for the “high software standards” required by malware corporatization, you and your company are going through some serious crap. You don’t even care how your company looks anymore - as long as it works.
But it won’t. And it never does.
I’m sorry, but this will probably be dead until either Unity’s CEO comes to his senses, or (more likely) things fail badly and he is eventually replaced. There is a reason EA won the golden swirl trophy under his tenure in 2012 (for being the worst company in America).
Does anyone know a good alternative to ML-Agents for quickly implementing reinforcement learning in a game? I started using Unity, and ML-Agents, because I was hoping to make a game with machine learning-based AI as a core mechanic. If Unity is abandoning it, then I should switch platforms. Any recommendations?
Two of the individuals giving that presentation at SIGGRAPH
were Senior ML developers at Unity. According to LinkedIn, they are both still employed at Unity. Sliver of hope.
Just found this thread, and it seems f*cked. Oo man, I was about to start learning ML agent, but its kind of risk for the long-term learning investment
“Well done to our ML Artistry team who won the Audience Choice Award at ACM SIGGRAPH 2022 Real-Time Live! Presented by Machine Learning Manager, Florent Bocquelet, “AI and Physics Assisted Character Pose Authoring” won the audience vote for their new research project around character animation and posing using machine learning. Unity’s character pose tool showed how users can work with complex, handcrafted constraints even with minimal artistic experience. We’re very proud of the team for this accomplishment, and congratulations to all the other speakers for their fascinating presentations”
So this is most likely the part that is left from A.I. stuff. And these ML agents is R.I.P
For me it was the perfect way to get started in reinforcement learning. Because of the easiness of setup and well documented hyperparameter tuning. And the huge amount of examples of course.
When it comes to reward assignment, what I think is most important part of RL, you wont risk much. Same strategies you can apply in any framework. Even if you go straight to Stable Baselines, etc.
And the current product is still pretty stable. You can create some vision of yours for sure. Maybe that is what this project needs in order to be resurrected → attention - projects build on it - preferably commercial, so cash flows in to Unity
Here’s a related paper by some of the same authors https://arxiv.org/pdf/2208.08274.pdf
none of whom I recognize as being part of the previous ML-Agents team.
It’s an interesting idea: estimating character poses from reference images and generating ml-aided IK for them. I can see how this could be rolled out as a productivity tool. However, it doesn’t really look like a reinforcement learning problem you would solve with ML-Agents, and I don’t expect there to be much overlap with the latter.
Besides, Nvidia seems to be ahead when it comes to creating physics-based character animations with reinforcement learning:
