Runtime launched at YC with a robot that plans in a copy of itself

Most robots that start a new task first need many hours of a person driving them by hand. Runtime launched at Y Combinator this week with a different method. Its software builds a copy of the world and plans the task inside that copy first. The company's launch page on Y Combinator has 55 upvotes.

The system turns what the robot sees into a simulation. It tries several actions inside that simulation. It picks the one that works, does it in the real world, then updates the simulation to match what happened.

What the company does

Runtime describes itself as a simulation and evaluation platform for physical AI. The product has two parts. One is the policy itself. The robot plans and acts without being trained on the task first and without a person steering it. The other is a test bed. A team builds a scene there, then runs the policy across thousands of variations of it. The team sees where the policy breaks before the robot is on a real site.

The founders are Liam, Yim, and Chen Peng, all from the GRASP lab at the University of Pennsylvania. The GRASP lab is one of the main robotics groups in United States academia. The three have published and reviewed work at CVPR, ECCV and NeurIPS.

Runtime is backed by the NVIDIA Inception Program and Pareto. NVIDIA Inception is NVIDIA's startup support programme and gives credits for cloud compute. The company says early access to the product opens in January.

How the simulation step works

Robot policies today come in two shapes. The first is trained on many hours of a person moving the robot through the task. This is called teleoperation. The policy is then fine-tuned. The second is a general policy from a frontier model that is given the task as a prompt. Both can fail on a task the robot has not seen before.

Runtime is a third shape. The robot captures the scene, builds a working model of it in software, and runs the policy against that model. That is the planning step. The robot then acts on the plan. It watches what happens. When the software model and the real world do not match, it corrects the software model.

Runtime's own description names rope threading and picking up an egg as example tasks. Both are tasks where the robot has to care about contact force without crushing the object.

What we can and cannot say

Runtime says its early results match or beat the leading robot foundation models on real-world tasks. That claim is the company's own. There is no named customer in the launch post, no published benchmark score, no disclosed round amount, and no price.

Y Combinator's launch page lists the company in Industrials, with the tags Robotics, Infrastructure and AI. The product opens in January. Before then, the next thing to ask Runtime for is a video. Or a score on a public benchmark that someone outside the company can reproduce. Until then, 55 YC votes and the credentials of three researchers from GRASP are the signal.

Source

• Primary source: Y Combinator: Runtime launch post, by Liam, Yim and Chen Peng.
• Company site: runtime.si.