Blogs›Runway›Robotics Policy Model
Runway introduces Data Augmentation for its General World Model (GWM-1) in robotics. This feature allows engineers to generate photorealistic variations of existing robot trajectories, including new environments, lighting, camera angles, and object placements. This enables scaling training data without additional collection by transforming existing trajectory videos into new conditions or generating full trajectory videos from a single frame and text prompt. This complements Offline Policy Evaluation by providing a richer and more diverse dataset for training and testing robot policies.
9/24/2025
This post details advancements in accelerating robot policy evaluation using General World Models, demonstrating how GWM-1 can be leveraged to significantly speed up the process of evaluating and refining robot policies.
Unknown date
Introduces Data Augmentation capabilities for GWM-1, enabling the generation of diverse training data through photorealistic variations of existing robot trajectories. This includes video-to-video generation (changing environments, lighting, obstacles while preserving action sequences) and image-to-video generation (creating full trajectories from a single frame and text prompt). The goal is to expand the training distribution without collecting more physical data.
Unknown date
This post introduces the Offline Policy Evaluation SDK for Runway's GWM-1 robotics model. It details the capability to simulate policy rollouts by feeding action sequences and camera observations into the model. The core technical contribution is the ability to surface failures, evaluate edge cases under diverse scenarios and lighting, and enable rapid iteration on policies by running hundreds of simulated rollouts in minutes, thereby reducing the need for physical robot testing.
Unknown date
This post introduces the Runway Robotics Policy Model, a new capability that leverages a General World Model (GWM-1) to predict robot actions from multimodal sensory input. It details the process of fine-tuning the model on robot demonstration data, receiving a custom policy model optimized for specific embodiments and tasks, and deploying it via a Python SDK for real-time inference. The post includes a code snippet demonstrating SDK integration for policy creation and inference.