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

arXiv 2026 45.7 method, system

TLDR

Pelican-Sim 1.0 is a general world model simulator that predicts future observations from actions, using unified action representation, sparse MoE, and efficient rollout; it shows strong benchmark gains.

Reasoning

The paper introduces a principled simulator architecture with clear innovations and validates them via extensive quantitative comparisons and downstream tasks on multiple benchmarks. However, as a technical report, the abstract reveals limited information about implementation details, baselines, and potential limitations, so some claims require fuller scrutiny.

Read-first score

Read-first score 45.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 62.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 25%

88.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,evaluation,experiment,metric

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=checkpoint

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 762.

Keyword Scores

world model

world simulator

interactive world model

video world model

world dynamics prediction

generative world model

model-based reinforcement learning world model

Deep Analysis

Innovations

General world model simulator that predicts future observations from visual context and robot actions for embodied intelligence
Unified 28-dimensional action value space covering mainstream embodiments, enabling one model across heterogeneous devices
Action-visual injection using URDF- and camera-rendered action videos to bridge actions and pixels, improving controllability
Sparse mixture-of-experts layers that add capacity for heterogeneous dynamics and reduce inter-modality conflict
Efficient rollout generation via causal adaptation and few-step distillation, yielding a four-step autoregressive simulator
Training on approximately one million real-world and simulated trajectories

Methodology

Pelican-Sim 1.0 is a world model simulator trained on about one million real-world and simulated trajectories to predict future observations from visual context and robot actions. It combines a unified 28-dimensional action representation, URDF- and camera-rendered action videos for action-visual injection, sparse mixture-of-experts layers, and causal adaptation with few-step distillation to create a four-step autoregressive simulator. Evaluation covers video quality and controllability metrics against dense and alternative fusion baselines on AgiBotWorld Beta, RoboMIND, and RoboTwin, plus downstream policy experiments on RoboTwin.

Key Results

Pelican-Sim 1.0 improves PSNR over the strongest evaluated baselines by 4.636 on AgiBotWorld Beta, 2.080 on RoboMIND, and 10.343 on RoboTwin, and raises adapted EWMBench DYN score by 0.426 on RoboTwin. Downstream on RoboTwin, adding 500 generated trajectories to 50 demonstrations per task raises policy success from 70% to 93%, policy evaluation reaches Pearson correlation of 0.994 across five checkpoints, and relative success gains reach 47.7% for action selection and 20.3% for policy improvement.

Resources

Paper