
Runtime is building a simulation-based robot policy that needs no fine-tuning or teleoperation.
https://youtu.be/FgQBFk13k6M
More companies than ever are deploying robots. For them, deployment should be plug-and-play: show the robot the task, and it works out the rest.
Runtime turns what the robot sees into a simulation, searches through options, acts on what works, and keeps the simulation in sync with reality. In our experiments, Runtime matches or beats leading robot foundation models on real-world tasks like rope threading and egg grasping. Next, we’re speeding up the loop so it can run live.
We’re Liam, Yim, and Chen from Penn GRASP. We’ve published/reviewed at CVPR, ECCV, and NeurIPS, with experiences in CV, control, RL, and robot foundation models. We’ve seen simulation become fast and accurate enough for more than training and evaluation, and we’re bringing sim directly into how robots act.
If you’re building or deploying robots, we’d love to hear what yours struggle with. Early access opens in January at runtime.si.