The robotics firm has released the code for its internal physics solver, Steel, to help developers create more efficient virtual training environments for autonomous machines. Developed in partnership with Nvidia, the solver reduces GPU memory usage by approximately 3x and performs simulations at speeds up to 13x faster than those of the industry-standard PhysX.
The project is the first open-source release from Skild AI and targets the ‘sim2real gap,’ the difficulty of transferring robot behaviors from a simulation to the real world. The company stated that Steel’s ability to handle heterogeneous environments at scale will allow coding agents to generate the diverse scenarios necessary to train robots for complex manipulation tasks.
