Training an AI spider in Unity with MLโAgents, CUDA, and PyTorch. Watching it learn to crawl from scratch never gets old. ๐ท๏ธโ๏ธ
#AI #MachineLearning #Unity3D #MLAgents #GameDev #AIDev #ReinforcementLearning #IndieDev #solodev #ai #livedev
www.youtube.com/live/REznci_...
AutoPlay Enables Scalable Synthetic Task Generation for AI Agents
AutoPlay generated 20โฏ000 tasks for 20 Android apps and 10โฏ000 tasks for 13 Ubuntu apps, boosting agent success rates up to 20.0โฏ% on mobile and 10.9โฏ% on desktop. getnews.me/autoplay-enables-scalabl... #autoplay #synthetictasks #mlagents
Evaluation-Centric Approach for Scientific Visualization Agents
Multimodal LLM breakthroughs have produced agents that create visualizations, but benchmarks are missing. The authors propose a community benchmark, published September 2025. Read more: getnews.me/evaluation-centric-appro... #sciviz #mlagents
After many years at Unity, working on a range of projects from #MLAgents and #ReinforcementLearning, to building agents for Muse Chat/Assistant to help developers make games faster, I needed a change. My goal has always been to build generally capable agents that can operate in the real world.
Me๐ vs a ML agent โค๏ธ in a pong match. This was a little fun project that allowed me to experiment with the ML agents Unity Package.
The training lasted about 15 minutes in self-play mode.
#unity #gamedev #indiedev #ai #mlagents
How the Perrinn 424 Agents "sees' it's environment... in addition to these exteroceptive sensors, the agent also has proprioceptive sensing of velocity, acceleration, etc etc, and some interoceptive sensing as well. #mlagents #unity #perrinn424project
Making progress... trained with ML-Agents Reinforcement Learning toolkit in Unity. #mlagents #unity #perrinn424project
Finally got the this working with the physically simulated character. It was tough getting the ground acceleration observations working properly. #madewithunity #mlagents
Latest training result for my 30 meter troll. Tried punishing it for dragging its feet while stepping. This is what it learned to do: #madewithunity #mlagents
Staggering when trying to balance on moving platforms. Powered ragdoll trained to walk using machine learning. No animations. #mlagents
Can now layer animations on top of the ML movement. The troll is copying an animated throw with the arms while walking with the legs. #mlagents #madewithunity
Experimenting with querying the ML Models less frequently... I think I need more damping. Feels a little loose. #mlagents
Using different models (Jump, Tumble) to create initial conditions for the Walk training. I love watching it improve over time... Really feel for the guy #mlagents #madewithunity
Started training a giant colossus size boss character. First steps (partially trained walk).
#unity3d #mlagents
Back from summer break. Next version of my physically simulated character controller. Faster, can walk backwards, more consistent jumps, more stable.#unity3d #mlagents
youtu.be/5QiMe6hTrnI
Added a stabilizer that keeps his center of mass over his feet while airborne. Not physically realistic but it makes the jumps much more reliable.
#madewithunity #mlagents
Added a heightfield-camera feature that measures the terrain so the dude can be trained to walk on rough surfaces and step over gaps. The green blobs are his foot positions. The red tint is how far the terrain is below the hips. #madewithunity #mlagents
I trained a tumbling brain for my physically simulated walking dude. #mlagents #madewithunity
Finally solved an ML-Agents problem since updating to v0.10 a month ago. Restart would have a huge drop in reward that took a long time to recover from. Discovered that I needed to add a Sampler file. Without it, initial conditions never changed. Severe over-fitting. #MLAgents
Combined run brain, jump brain, getup brain to drive fully controllable physically simulated character controller. Starting to feel usable but not very robust yet. #mlagents #MachineLearning
First attempt at integrating multiple brains to get up from prone and supine fall positions. #mlagents.
Added a reward for matching a pose and got a much better result. Although I accidentally set the target height too high. To get the reward it is standing on its tippy toes. #mlagents
Decided to try ML to animate the bug. First attempt, try to get it to stand up. Hmm, needs some work. Its belly is not touching the ground though. #mlagents
