Runtime, a startup with a Y Combinator launch post, has launched a robot policy that simulates each action before the machine carries it out, with no fine-tuning or teleoperation, according to the company's YC launch post. It has not published success rates, and a crowded field is chasing the same problem: the time and labor needed to adapt a robot to each new task.
How the policy works
The YC launch describes a four-step loop. The system reconstructs the scene it observes, searches for candidate actions in simulation, executes one, and then synchronizes the simulation with what actually happened in the physical world.
Here, simulation is used to make decisions at run time. It is not merely a source of synthetic training demonstrations, which is one common way robotics teams employ it.
The post names rope threading and egg grasping as example tasks. It offers no numerical success rates, trial counts, named comparison models or reproducible evaluation protocol. It also acknowledges that work remains to speed up the loop enough for live operation. Performance, in other words, is the company's own claim, and production-speed deployment has not been established.
Nor is this a financing or customer announcement. The company's website directs visitors to an early-access signup.
Who is behind it
Chen Peng is co-founder and chief executive, according to the launch post. Ningze "Liam" Zhong is co-founder and chief technology officer, per his personal website, which carries an October 2026 news item announcing Runtime's co-founding.
That site also mentions Pareto Holdings, YC and NVIDIA Inception, a program for startups. It specifies no investment amounts, dates or terms, and no company financing announcement turned up in the research for this article.
"No fine-tuning" is a crowded claim
Runtime is hardly alone in skipping parameter updates, the retraining step that adjusts a model's internal weights. Simify, a paper on the arXiv preprint server dated September 27, 2026, describes training-free, test-time planning through parallel physics simulation. Like Runtime, it places simulation inside the selection of actions.
Generalist's GEN-1.5 announcement, from August 2026, describes adaptation through in-context demonstrations without parameter updates. A demonstration may still be needed even when fine-tuning of model weights is not.
NVIDIA's GR00T 2, announced in November 2026, reports zero-shot capabilities (performing tasks without task-specific training), but it relies on substantial human-video and robot-teleoperation data during pretraining. Runtime has not disclosed a comparable inventory of training data. "No teleoperation" and "no human-derived data" are different claims, and the company's post addresses only the first.
Physical Intelligence took another route. Its October API announcement packages the existing adaptation loop as managed infrastructure, which lightens the burden of fine-tuning without eliminating it. The company also acknowledged connectivity problems in some settings. A separate academic project, LEGS, aims to remove teleoperation data collection while keeping fine-tuning.
The market Runtime is chasing
The International Federation of Robotics (IFR) released its World Robotics 2026 report on September 24, with its president, Jane Heffner. The press release was headlined "Five million robots now operate in factories globally." The federation expects easier programming, machine vision, sensing and AI to widen the range of economically feasible applications.
Those figures describe industrial robots in general. They say nothing about Runtime or about simulation-based policies.
The IFR's January 2026 outlook says buyers weigh cycle time, energy consumption, maintenance, safety, durability and consistent performance, not intelligence alone. Deepak Pathak and Abhinav Gupta, founders of Skild AI, wrote in a September 10, 2026 post that installation, integration, maintenance and process updates are core problems.
Pricing models vary across the sector. The IFR describes robots-as-a-service as the dominant US model for transport and logistics robots. Runtime's own pricing and revenue remain unverified.
Proof is the gap
New evaluations aim to separate polished demos from repeatable performance. VABAR, presented in November 2026 materials from UC Berkeley, spans six tasks with standardized variations and success metrics. DexBench, launched November 9, 2026, covers dexterity across retail, logistics, manufacturing and hospitality.
Neither has been applied to Runtime, based on the research. Transferring skills from simulation to the real world is also decades-old research territory, so Runtime's novelty rests on its specific action-selection loop rather than on the idea of simulation itself.
The research located no reproducible Runtime benchmark and no independently measured production deployment. Runtime's October benchmarking article could not be retrieved in full.
For now, the next milestone is the one the company itself named: speeding up the simulation loop enough for live operation.
