RISE Framework Lets Robots Self-Improve in Imaginary Worlds, Hits 95%

Modern Vision-Language-Action (VLA) models excel at imitation learning but fail at contact-rich tasks like dynamic sorting or flexible packing, where small execution errors cause failures. Real-world reinforcement learning remains impractical due to high hardware costs, manual environment resets, and safety risks. To address this, researchers from CUHK MMLab, HKU OpenDriveLab, and Tsinghua University introduced RISE (RSS 2026), a framework that enables robots to self-improve entirely within simulated 'imaginary space.' RISE replaces expensive physical trial-and-error with a compositional world model that predicts future robot interactions and manipulates multi-view trajectories in seconds using a video diffusion model (Genie Envisioner). A progress value model built on the π0.5 VLA framework then evaluates behaviors via real-time advantage scores, detecting subtle failures like object slipping. This closed-loop pipeline allows a robot policy to warm up on small offline data, generate imagined rollouts, and optimize itself iteratively.

Using the AgileX Piper 6-DoF robotic arm, RISE achieved a 95% success rate on complex tasks such as deformable material handling and bimanual coordination. The dynamics model was pretrained on large-scale robot datasets (Agibot World, Galaxea) to ensure realistic predictions. By eliminating the need for physical resets and hardware wear, RISE dramatically lowers the barrier to scaling robot learning. It represents a major step toward general-purpose embodied intelligence, where robots can autonomously refine their skills in simulation before deploying in the real world. The project is open-source on GitHub, and the paper is available on arXiv.