paper/arxiv/2608.29078

Robotics Dataset

This paper introduces DREAM, a framework that generates fine-tuning data for vision-language-action (VLA) models in new workspaces without human demonstrations, using real-to-sim reconstruction, LLM-based task planning, and task-and-motion planning to create synthetic trajectories and image-action examples.

Catalog metadata

Repository owner

Makoto Sato, Tatsuya Matsushima, Yutaka Matsuo et al.

Source

Research paper

Task

DREAM: Deployment-Time Demonstration Generation via Real-to-Sim for Scalable Policy Adaptation

Cameras

0

Browse the catalog to compare robots, tasks, sensors, and licenses. Sign in to open protected download and model links.