Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence
Shilong Zou, Shilin Zhang, Yingji Zhang, Yuhang Huang, Yi Zhang, Zeyuan Ding, Han Dong, Junwei Liao, et al.
cs.RO · cs.AI
Quick verdict
Claims large gains for a cross-embodiment world simulator; strong ablations, but no error bars and possible train-eval overlap temper the verdict.
Summary
This technical report from X-Humanoid presents Pelican-Sim 1.0, an action-conditioned video world model for robot manipulation. It conditions a 28-layer Video DiT (built on Cosmos-Predict 2.5) on two synchronized action representations: a unified 28-dimensional configuration vector covering single/bimanual arms with grippers or dexterous hands, and URDF-rendered, camera-aligned skeleton videos. The two modalities are injected into alternating blocks (scale-shift modulation vs. context residuals), sparse MoE layers (1 shared, 8 routed experts, top-2 sigmoid routing) replace dense FFNs, and progressive causal distillation compresses 35 denoising steps to 4 (5.67x speedup, 2.2 s per 21-frame clip). Trained on roughly one million real and simulated trajectories (8,000 hours), it reports the best PSNR/SSIM/LPIPS/FID/FVD on held-out splits of AgiBotWorld Beta, RoboMIND, and RoboTwin (e.g., PSNR 30.383 and FVD 4.980 on RoboTwin vs 20.040/26.600 for Ctrl-World). Four RoboTwin applications follow: 500 generated trajectories on 50 demonstrations raise pi0.5 success from 70% to 93%; a fine-tuned Qwen3-VL evaluator yields Pearson r=0.994 across five checkpoints (1,000 adaptation rollouts); action selection gains +20.6 points (43.2% to 63.8%); GRPO policy improvement reaches 75.4% from a 62.7% SFT start. Limitations: no confidence intervals or significance tests, within-source splits with RoboTwin appearing in both training and downstream evaluation, qualitative-only OOD evidence, and some EWMBench sub-metrics still below the best baselines.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 0 weaknesses · 0 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Pairing a unified 28-D action vector with URDF-rendered, camera-aligned skeleton videos injected via alternating pathways improves action-controllable video quality over either modality alone (+1.000 dB PSNR over action-video-only on AgiBotWorld Beta).
- Replacing dense FFNs with token-choice sparse MoE layers under mixed-domain training reduces FVD by 6.530 vs the dense backbone on AgiBotWorld Beta (Table 4).
- Causal adaptation plus few-step distillation yields a 4-step autoregressive simulator with a measured 5.67x speedup (2.2 s per 21-frame clip), enabling rollout-intensive downstream use.
- Generated rollouts are directly useful: 500 generated trajectories on 50 demonstrations lift pi0.5 success from 70% to 93% on RoboTwin; the world-model+VLM pipeline ranks five checkpoints consistently (Spearman 1.0, Pearson 0.994 at 1,000 adaptation rollouts); action selection gains +20.6 pp and GRPO policy improvement gains +12.7 pp over SFT.
Strengths
Comprehensive, cumulative ablation suite
Tables 4-6 cleanly separate the factors: data mix and MoE (PSNR 20.514 to 21.062 to 22.276; FVD 20.410 to 17.380 to 10.850 on AgiBotWorld Beta), action modality contribution (action-video-only 21.276 vs action-value-only 19.238 PSNR), and injection architecture (interleaved within 0.075 dB of dual-every-block, justifying the cheaper design)
Consistent large video-quality gains against retrained baselines
Baselines (IRASim, Ctrl-World, EnerVerse-AC, OSCAR, Wan-Move) are retrained on the same mixed-domain corpus (Section 4.2); gains are large: RoboTwin PSNR 30.383 vs 20.040 and FVD 4.980 vs 26.600 for the strongest baseline
Downstream utility shown end-to-end with oracle controls
Candid self-limitation
Concrete, measured efficiency results
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Pelican-Sim 1.0 (arXiv 2609.12036) is a 27-page industrial technical report from X-Humanoid's WFM group proposing an action-conditioned video world model for robot manipulation. Four contributions are claimed: (1) a unified 28-dimensional action interface (7 arm joints + 1 gripper + 6 hand joints per side) spanning single/bimanual arms; (2) complementary action-video conditioning - URDF forward-kinematics skeleton renders, camera-aligned per frame - interleaved with action-value modulation in a 28-layer Video DiT built on Cosmos-Predict 2.5; (3) token-choice sparse MoE layers (1 shared + 8 routed experts, top-2 sigmoid routing) for heterogeneous dynamics; (4) a 4-step causal distilled simulator (5.67x faster than the 35-step teacher, 2.2 s per 21-frame clip). Downstream claims: best PSNR/SSIM/LPIPS/FID/FVD on held-out splits of three benchmarks, plus four RoboTwin applications - demonstration augmentation (70% to 93% pi0.5 success), VLM-assisted policy evaluation (Pearson r=0.994 over five checkpoints), action selection (+20.6 pp to 63.8%), and GRPO policy improvement (62.7% to 75.4%).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 0 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Pelican-Sim 1.0?
It is an action-conditioned video world model from X-Humanoid's humanoid robotics center. Given one RGB frame, a unified 28-dimensional robot action sequence, and a URDF-rendered camera-aligned skeleton "action video", a 28-layer Video Diffusion Transformer with sparse Mixture-of-Experts layers predicts the resulting future camera observations, and a 4-step causal distilled version generates rollouts fast enough for downstream policy training, evaluation, and action selection.
What results does the paper report?
On held-out splits of AgiBotWorld Beta, RoboMIND, and RoboTwin it reports the best PSNR/SSIM/LPIPS/FID/FVD among retrained baselines (RoboTwin: PSNR 30.383, FVD 4.980 vs 20.040/26.600 for Ctrl-World). On RoboTwin downstream tasks: 500 generated trajectories added to 50 demonstrations raise pi0.5 success from 70% to 93%; a fine-tuned VLM evaluator reaches Pearson r=0.994 across five policy checkpoints; action selection gains +20.6 points (43.2% to 63.8%); GRPO policy improvement reaches 75.4% from a 62.7% SFT start.
Is the code or model available?
The paper promises to release model checkpoints and inference code (GitHub: Open-X-Humanoid/Pelican-Sim1.0), but as of the report date they are not yet available. Full retraining used about one million trajectories and 32 NVIDIA B200 GPUs, which is impractical outside large labs; downstream protocols, however, use public assets (RoboTwin simulator, pi0.5, OpenVLA-OFT, Diffusion Policy, Qwen3-VL-2B).
What are the main criticisms of the paper?
The main criticisms are: no confidence intervals or significance tests anywhere (the headline r=0.994 rests on only five checkpoints); RoboTwin appears both in the training corpus and in every downstream evaluation, so domain overlap cannot be excluded; OOD generalization across embodiment, viewpoint, scene, and object is shown only with qualitative figures; and a few adapted EWMBench sub-metrics (e.g., SceneC, Diversity, some RoboMIND motion scores) remain below the best baselines.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure contents (Figs. 1-11) are not extractable from the PDF text; the qualitative generation-quality and five-axis OOD claims can only be assessed from captions and the surrounding prose, not from the images themselves.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.