Researchers have released Pelican-Sim 1.0, described as a general world model simulator for embodied intelligence. According to the technical report, the system predicts future observations from visual context and robot actions in order to support downstream learning and decision making.
The model rests on four design features. A unified action representation uses a 28-dimensional action value space intended to cover most mainstream embodiments, so a single model remains valid across heterogeneous devices. Action-visual injection uses URDF- and camera-rendered action videos to bridge actions and pixels, which the authors report improves controllability across embodiments, scenes, and tasks, with a PSNR gain of 0.904 over alternative fusion baselines.
A sparse mixture-of-experts (MoE) design adds capacity for heterogeneous dynamics and absorbs the action modality while reducing inter-modality conflict, yielding an FVD improvement of 6.530 versus the dense backbone. For efficient rollout generation, causal adaptation and few-step distillation produce a four-step autoregressive simulator that achieves a 5.67-fold speedup over the 35-step model.
Trained on approximately one million real-world and simulated trajectories, Pelican-Sim 1.0 reports PSNR improvements over the strongest evaluated baselines of 4.636 on AgiBotWorld Beta, 2.080 on RoboMIND, and 10.343 on RoboTwin, with the adapted EWMBench DYN score up 0.426 on RoboTwin.
The report also describes four downstream applications on RoboTwin: adding 500 generated trajectories to 50 demonstrations per task raised policy success from 70% to 93%; policy evaluation reached a Pearson correlation of 0.994 across five checkpoints; and relative success gains reached 47.7% for action selection and 20.3% for policy improvement. Qualitative generalization is reported across trajectory, scene, object, embodiment, and viewpoint shifts.