Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence
Pelican-Sim 1.0 is pitched as a robot world-model simulator that can roll out future visual observations from action inputs across different embodiments.
The report says the model uses a unified 28-dimensional action space, action-video injection from URDF and camera renders, sparse MoE layers, and a faster four-step rollout setup. Trained on about one million real and simulated trajectories, it beats evaluated baselines on PSNR across AgiBotWorld Beta, RoboMIND, and RoboTwin. On RoboTwin, generated trajectories improved policy success from 70% to 93% when added to 50 demonstrations per task, and policy evaluation reached a 0.994 Pearson correlation across five checkpoints. HF Daily Papers' note
The report says the model uses a unified 28-dimensional action space, action-video injection from URDF and camera renders, sparse MoE layers, and a faster four-step rollout setup. Trained on about one million real and simulated trajectories, it beats evaluated baselines on PSNR across AgiBotWorld Beta, RoboMIND, and RoboTwin. On RoboTwin, generated trajectories improved policy success from 70% to 93% when added to 50 demonstrations per task, and policy evaluation reached a 0.994 Pearson correlation across five checkpoints. HF Daily Papers' note
score 6