Pelican-Unify 1.0: A Unified Embodied Intelligence Model for Understanding, Reasoning, Imagination and Action

arXiv 2026 39.6 method

TLDR

Pelican-Unify 1.0 unifies vision-language understanding, reasoning, future video/action generation, and action in one embodied model, showing strong benchmark results.

Reasoning

The paper proposes a unified architecture with shared VLM and latent conditioning for future video/action generation, supported by benchmark results across understanding, video generation, and robot action. However, the abstract lacks details on baselines, limitations, and real-world deployment, making some claims hard to fully assess.

Read-first score

Read-first score 39.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 45.

Recency 5%

100

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

Methodology quality 15%

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

Topical relevance 25%

64.3

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

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: 519.

Keyword Scores

generative world model

video world model

world model

world dynamics prediction

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces Pelican-Unify 1.0 as a unified embodied foundation model that uses a single VLM to map scenes, instructions, visual contexts, and action histories into a shared semantic space.
Uses the same VLM as a unified reasoning module that autoregressively produces task-, action-, and future-oriented chains of thought in one forward pass and projects the final hidden state into a dense latent variable.
Proposes a Unified Future Generator (UFG) that conditions on the latent variable and jointly generates future videos and future actions through two modality-specific output heads within the same denoising process.
Jointly backpropagates language, video, and action losses into the shared representation, enabling unified optimization of understanding, reasoning, imagination, and action instead of training isolated expert systems.

Methodology

Pelican-Unify 1.0 uses a single VLM for both understanding and reasoning, mapping diverse inputs into a shared semantic space and autoregressively outputting chains of thought plus a dense latent representation. A Unified Future Generator conditions on this latent variable and jointly generates future video and action outputs via two modality-specific heads within one denoising process. Language, video, and action losses backpropagate into the shared representation for joint optimization; training data details are not specified in the abstract.

Key Results

With a single checkpoint, Pelican-Unify 1.0 reports 64.7 on eight VLM benchmarks (best among comparable-scale models), 66.03 on WorldArena (first), and 93.5 on RoboTwin (second-best average among compared action methods). The authors interpret these results as showing the unified paradigm preserves specialist strength.

Limitations

No explicit limitations are stated in the title or abstract.

Resources

Paper