Researchers have built Pelican-Unified 1.0, a single AI model that combines visual understanding, task reasoning, future imagination, and robot action into one cohesive system. This unified approach achieves state-of-the-art performance across vision-language, world modeling, and action benchmarks while enabling zero-shot transfer to real robots without additional training.
What the Researchers Built
Pelican-Unified 1.0 is a foundation model for embodied intelligence that integrates four interdependent capabilities—understanding (perceiving the current scene), reasoning (interpreting tasks), imagination (predicting future states), and action (executing physical movements)—into a single trainable neural network. Unlike previous approaches that chain separate vision models, language models, world models, and action policies together after training, Pelican-Unified learns a shared latent world space where all four functions co-evolve.
The model was trained on loop-closed data: examples that simultaneously contain observation, instruction, reasoning traces, action sequences, and future outcomes. This allows the model to absorb coupled signals rather than isolated tasks. The researchers evaluated the model on eight vision-language benchmarks, the WorldArena world-model benchmark, and the RoboTwin dual-arm simulator, as well as in real-world tests using a UR5e robotic arm and a Tienkung humanoid robot. The system performed zero-shot transfer—generalizing to new tasks and environments without any extra fine-tuning.
Key Results
Pelican-Unified 1.0 matches or exceeds specialist models across all three capability domains. On eight multimodal vision-language benchmarks, it scored an average of 64.7, the best among comparable-scale models. On WorldArena, a benchmark for action-conditioned world models, it achieved 66.03, ranking first overall. On the RoboTwin simulator, it reached an average score of 93.5, the second-best among compared action methods.
In real-world human evaluation with blind expert scrutiny, Pelican-Unified 1.0 achieved an overall mean score of 1.76 (rank #1), outperforming the best video-diffusion specialist (Seedance2.0, 1.55) by 0.21. Notably, it earned a perfect 2.00 in controllability and the highest task success score of 1.81. The model also demonstrated zero-shot generalization: performing compositional tasks in new environments without any additional training data. The unified training approach did not degrade specialist competence—instead, it preserved or improved each capability.
How It Works
Pelican-Unified 1.0 is built on a shared latent world space that aligns perception, reasoning, imagination, and action. The model processes a visual observation and a natural language instruction through a unified transformer backbone. It then generates a sequence of imagined future frames (the "imagination" phase) while simultaneously decoding motor commands (the "action" phase) conditioned on those imagined futures. This is not a two-stage pipeline but a joint decoding process where the model learns to produce both video predictions and action tokens from the same latent representation.
The training loop uses data that couples all four signals: an image, a task instruction, a reasoning trace, a sequence of future images, and corresponding robot joint commands. The model is trained end-to-end to minimize a combined loss over language, video, and action outputs. This forces the representations to be mutually informative—reasoning directly conditions generation, and future video is decoded jointly with action.
| Benchmark / Task | Pelican-Unified 1.0 | Best Comparable Specialist | Improvement |
|---|---|---|---|
| VLM Benchmarks (8 avg) | 64.7 | — (best among comparable-scale) | — |
| WorldArena (world model) | 66.03 | — (rank #1) | — |
| RoboTwin (action) | 93.5 | — (rank #2) | — |
| Human Evaluation (overall) | 1.76 | 1.55 (Seedance2.0) | +0.21 |
| Controllability (human eval) | 2.00 (perfect) | — | — |
| Task Success (human eval) | 1.81 | — | — |
The researchers also performed ablations showing that loop-closed data—where observation, instruction, reasoning, action, and future outcomes are annotated on the same example—provides the most valuable training signal. A unified model can absorb this coupled data directly, whereas modular systems would require separate training stages.
Why This Matters for Robotics
Pelican-Unified 1.0 challenges the current modular paradigm in embodied AI. Until now, most systems have stitched together separate vision, language, world, and policy models after training. This work shows that training a single model to jointly perform understanding, reasoning, imagination, and action yields stronger integrated behavior—especially in zero-shot transfer and compositional generalization—without sacrificing specialist performance.
For real-world applications, this means robots that can handle novel tasks and environments out of the box, using the same model for perception, planning, and control. This is particularly valuable for warehouse robots that need to adapt to changing layouts, and for humanoid robots that must reason about complex manipulation tasks. The ability to imagine future states and condition actions on those predictions could reduce trial-and-error learning, improve safety, and enable long-horizon tasks like assembly or cooking. Pelican-Unified demonstrates that a single checkpoint can replace a stack of specialists, simplifying deployment and maintenance for robotics engineers.
Limitations and Open Questions
The researchers acknowledge that Pelican-Unified 1.0 does not claim "general embodied intelligence." The model was evaluated on a limited set of benchmarks and real-world tasks; scaling to more diverse environments and longer horizons remains an open challenge. The real-robot experiments were conducted with only two platforms (UR5e and Tienkung), and it is unclear how well the model transfers to other robot morphologies. Additionally, the computational cost of training such a unified model is higher than training individual specialists—though the authors note that a single checkpoint is simpler to deploy. Future work will need to explore whether the unified paradigm scales to even larger models and broader task distributions, and whether the loop-closed data requirement can be satisfied at scale.
Frequently Asked Questions
Does Pelican-Unified 1.0 require separate models for each capability? No, it uses one shared model that learns understanding, reasoning, imagination, and action together in a single training process.
How does the model perform zero-shot transfer to real robots? The model generates imagined future frames during inference and decodes motor commands conditioned on those predictions—no additional fine-tuning is needed for new tasks or environments.
What hardware was used in real-world tests? The team used a UR5e robotic arm and a Tienkung humanoid robot for all real-world validation experiments.
Is Pelican-Unified 1.0 better than modular approaches? Yes, it achieves competitive or superior results on each specialist benchmark while also enabling behaviors like zero-shot transfer and compositional generalization that modular systems struggle to implement.
Conclusion
Pelican-Unified 1.0 demonstrates that a single model can integrate perception, reasoning, imagination, and action into one adaptive system without sacrificing specialist performance. This unified paradigm may reshape how the robotics community builds foundation models for embodied intelligence, moving away from assembling larger specialists toward learning a shared process.