Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence

Details

Zou et al. introduce Pelican-Sim 1.0, a world model simulator designed to bridge heterogeneous robot embodiments under a single architecture. Four design choices distinguish it: a 28-dimensional unified action space covering most mainstream robots; URDF- and camera-rendered action videos as an intermediate representation bridging actions and pixels (PSNR +0.904 over alternative fusion baselines); sparse Mixture-of-Experts layers to handle heterogeneous dynamics with less inter-modality conflict (FVD -6.530 vs. dense backbone); and a 4-step autoregressive inference path achieving 5.67× speedup over the 35-step model via few-step distillation. Trained on approximately one million real-world and simulated trajectories, Pelican-Sim achieves PSNR gains of 4.636 on AgiBotWorld Beta, 2.080 on RoboMIND, and 10.343 on RoboTwin over the strongest evaluated baselines. Four downstream applications validate practical utility: adding 500 generated trajectories to 50 demonstrations per task raises policy success from 70% to 93%; policy evaluation achieves a Pearson correlation of 0.994 across five checkpoints; and relative success gains reach 47.7% for action selection and 20.3% for policy improvement. The model generalizes across trajectory, scene, object, embodiment, and viewpoint shifts, making it a credible candidate for general-purpose robotic world modeling. No weights or code release is announced in this technical report.