The platform optimizes data pipelines for humanoid robot training.
Building simulators for AI -- not just graphics but fully optimized to produce good data
We are super excited to share with you our initial release of Lucky Engine. We are building a robotics engine from the ground up to be what we wished we could find in a simulator before
Many users expressed excitement about Lucky Robots launching the Lucky Engine Robotics Simulator because it sounds fun and they are proud to share it with requests for future features, while a few voiced strong disapproval.
Download: https://luckyrobots.com/
Robotics need precise time management for physics and rendering. Typical game engines, like unreal, unity etc, don’t have this (we tried). If your CPU spikes while you play, game stutters a bit, then the game continues. That means dropped frames. Dropped frames means corrupt data for AI.
Next, you want to have your robots interacting with that environment. You can use C# scripting built into the editor, or use our gRPC API to communicate with our simulator in Python or any language you want.
Then you need to be able to create an environment that your robot can train on, this usually takes weeks and probably isn't what you want to do as a researcher. Even then, making them physically accurate is a lot of work. With Lucky Engine, you can create any number of physically accurate objects, homes, rooms, kitchens, and drop your robot in there for it to train on. Domain randomization included.
We are super excited to share with you our initial release of Lucky Engine. We are building a robotics engine from the ground up to be what we wished we could find in a simulator before
When you need to produce millions or billions of domain randomized episodes, you can use Lucky Hub to offload that work to us, and see your dataset created overnight. Lucky Hub is now live, and cloud-scaled recording will be coming very soon.
We needed to create an engine that is optimized for robotics and AI, not for playing games. With Lucky, you can have any # of robots, any # of sensors, any # of cameras per scene, and each frame state can be accurately recorded frame by frame.
Mujoco is our primary physics engine, and we use Jolt as a secondary. You can use mujoco for manipulation and if objects are far in the distance, or for things that you don’t need to record states, but still want to be visually accurate, you can use jolt. We will be adding more physics engines soon, since the LE architecture supports multiple engines, liquid, deformables and beyond.
There is no limit to what you can create on Lucky Engine, since it’s built like a game engine. It produces LeRobot 3.0 data format ready for you to train any policy.
We are very proud to share this with you, please let us know what you think and want added or changed! We will be adding a lot more very soon: Warp, RL frameworks, expanding scene generation, can’t wait to share with you all!
Download: https://luckyrobots.com/
@chris_j_paxton sounds like fuuunnnn
@luckyrobots Horro
@luckyrobots Let’s put those lazy dudes in a bootcamp training! Good luck!
@chris_j_paxton the hard part is making the simulation feel real enough that the ai doesnt learn the quirks instead of the logic
what physics are you prioritizing?
The platform optimizes data pipelines for humanoid robot training.
Building simulators for AI -- not just graphics but fully optimized to produce good data
We are super excited to share with you our initial release of Lucky Engine. We are building a robotics engine from the ground up to be what we wished we could find in a simulator before
