Runtime
Runtime is a research lab building robotics simulation technologies that enable frontier AI models to learn and act in simulated environments before real-world deployment. Its Nomad harness connects models to simulated robots through a standardized control loop, while a critic agent iteratively refines task-specific SKILL.md files through verbal reinforcement learning. In testing across 40 LIBERO tasks, this approach raised native success rates from 29% to 53.9%.
- Robotics
- Artificial Intelligence
- AI Agents
- Developer Tools
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
Frontier AI models can control robots, but their raw performance in manipulation tasks is often poor without task-specific guidance, achieving only 29% native success on standard benchmarks. Developing robotic skills through physical trials is slow, costly, and difficult to iterate on, limiting progress in agentic robotics.
Solution
Runtime is a research lab building robotics simulation technologies that enable agents to learn and act in simulated environments before deployment in the real world. The company's Nomad harness connects frontier models to simulated robots through a common control loop, allowing models to receive camera images, robot state, and feedback, then propose bounded actions. A critic agent reviews failed traces and writes SKILL.md task files that describe grasps, sequence steps, and failure recoveries, which are iteratively refined through verbal reinforcement learning without changing model weights. In testing across 40 LIBERO tasks, this skill-based approach raised native success rates from 29% without skills to 53.9% with best-observed skill cohorts, with gains demonstrated across multiple frontier models.
Target Audience
Primary customers are AI research labs, robotics developers, and enterprises building agentic robotic systems that need simulation-first training and evaluation before physical deployment.
Features
- Nomad harness providing a standardized action–observation control loop between AI models and simulated robots
- Task skill system using SKILL.md text files to pass task-specific knowledge like grasps, step sequences, and failure recovery procedures
- Verbal reinforcement learning loop where a critic agent reviews failed traces and iteratively refines skill files
- Multi-model support tested with GPT-6 Astra, Gemini 4 Argon, and Fable across 40 LIBERO benchmark tasks
- Simulation-based evaluation with recorded actions, outcomes, wall time, token usage, and video replay
- Model-agnostic architecture allowing models, task instructions, and simulators to change while maintaining consistent control contracts