Today we’re open-sourcing Steel, our internal physics solver built with @NVIDIA.
It runs simulations up to 13× faster than PhysX while using ~3× less GPU memory.
Steel unlocks scaleable simulation for manipulation, for everyone.
Today we’re open-sourcing Steel, our internal physics solver built with @NVIDIA.
It runs simulations up to 13× faster than PhysX while using ~3× less GPU memory.
Steel unlocks scaleable simulation for manipulation, for everyone.
A quick Fable 5.2 + Steel 0.1 test
“Make Wuji hand climb a spiral staircase by walking on its fingers as legs”
I feel like I am not ambitious enough 😁
https://t.co/U7lXV11vgE
We just built a rigid body solver that scales!
Simulation is a crucial component of robotic infrastructure, and I believe we should build it in open
Check out the blog, give it a try in Newton and send PRs my way, it’s time to accelerate🏎️💨
Scalable simulation for robotics is here.
Coding agents can now generate environments at scale. The next bottleneck? Simulating them at scale.
Introducing Steel: a scalable physics solver for heterogeneous environments. Our first open-source release at @SkildAI, built in partnership with @NVIDIA.
Our ambition is to rally the industry around closing the sim2real gap. Contributions welcome from you, your organizations, and your agents! ❤️
Really excited to finally introduce Steel, now fully open source in Newton!
Simulation infrastructure will be crucial to unlocking the scale of diverse experience needed for truly general robot manipulation. Building it in the open means anyone can try applying it to their own scenarios and as a collective everyone can push it forward.
Come build with us! We’re improving it every day, have already adopted it internally, and we’re pushing for Steel to become the go-to solver for large-scale RL.
https://t.co/p4S3lACEg2
Perf-wise, in lower DoF scenarios MuJoCo Warp is quite competitive, and in higher DoF cases Steel can win by large margins. We saw about 30% improvement across the board on IsaacLab tasks.
Dynamics-wise, Steel is closer to PhysX. In my experience, Steel is more forgiving and well suited to sim2real transfer via large-scale RL, which has always been a motivation for us. More on this later.
Happy to help try to port any projects/scenarios you might have in mind
@wan73184085@nvidia We include more comparisons in the blog, and observe a ~30% speedup on IsaacLab tasks over MuJoCo Warp!
Blog: https://t.co/gsIRRYC5L2
More details: https://t.co/g5Dx72IPPe
Simulating heterogeneous environments/tasks/robots at scale was a big motivation of this effort to make sure we design for the scale of diverse experience that learning general manipulation policies need.
MuJoCo Warp (integrated in Newton, just as Steel is) does not support heterogeneity in a natural way. The architecture builds a single MuJoCo model from a template world and runs parallel copies of it, meaning every world must have:
- the same number of bodies, joints, equality constraints
- matching entity types and per-joint DoFs
- matching actuator layouts
Every kernel in MuJoCo Warp assumes per-world data is organized as (worlds x fixed mjModel size), meaning you can vary parameters of a model, but not structure.
In contrast, we specifically designed Steel such that each world can have its own articulations and constraints, making it very suitable for heterogeneity.
We do include a comparison against MJWarp in the blog on simpler IsaacLab tasks and observe roughly 30% perf improvement.
To unlock that scale, we are actively encouraging and accepting contributions ❤️ and are committed to maintaining Steel: a shared foundation for diverse, scaleable simulation.
Code: https://t.co/rheJuPM7jq
Blog: https://t.co/gsIRRYC5L2
In our S1 blog post, we described simulation as scalable but limited in diversity.
Steel’s goal is to fulfill the full potential of diverse, scaleable simulation for robotic learning.