AnswersPublished17 sources
X-Humanoid’s Pelican Models Point Toward Unified Robot Intelligence
X Humanoid’s Pelican reports show complementary approaches to robot action, world simulation and embodied reasoning. Pelican VLA 0.5’s report integrates vision language understanding, future frame generation and action prediction; it reports 91.4% success on RoboTwin Clean and 91.0% on RoboTwin Randomized after fine...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- X Humanoid’s Pelican reports show complementary approaches to robot action, world simulation and embodied reasoning.
- Pelican VLA 0.5’s report integrates vision language understanding, future frame generation and action prediction; it reports 91.4% success on RoboTwin Clean and 91.0% on RoboTwin Randomized after fine tuning.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What roadmap did the Beijing Innovation Center of Humanoid Robotics (X-Humanoid) present through world model lead Zou Shilong’s “Exploring U. Article summary: The available evidence supports X-Humanoid’s broader direction: combining visual-language understanding, imagined future observations, and action generation in unified embodied models, with a world-model simulator suppor. Topic tags: general, general web, user generated, academic. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, char
X-Humanoid’s public Pelican reports point toward embodied models that connect perception, prediction and action. They describe three related but distinct roles: Pelican-VLA 0.5 generates actions alongside future frames, Pelican-Sim 1.0 predicts future observations conditioned on robot actions, and Pelican-Unify brings understanding, reasoning, imagination and action into a unified model. 17
18
44 The available sources do not verify that this full research direction—or the specific generalization taxonomy and self-evolving policy loop sometimes associated with it—was presented in Zou Shilong’s October ROSCon China keynote.
What the public reports establish
Pelican-VLA 0.5: connect perception to action
Pelican-VLA 0.5 is described as a unified vision-language-action model that combines vision-language understanding, future-frame generation and action prediction in one architecture. 44 Its project repository identifies a shared Qwen3-VL 4B backbone and a cross-embodiment pretrained checkpoint; the repository reports training on about 2,400 hours of heterogeneous manipulation data.
46
47
The report also describes an attention result: before task-specific fine-tuning, the model’s action pathway focuses on the object and contact region relevant to the instruction. 44
46 After fine-tuning on RoboTwin, the report gives success rates of 91.4% on RoboTwin Clean and 91.0% on RoboTwin Randomized.
44
Those findings concern Pelican-VLA 0.5. They do not verify the separate claim that adding 500 generated trajectories to 50 original demonstrations per task produced particular results for Pelican-Sim.
Pelican-Sim 1.0: predict what happens after an action
Pelican-Sim 1.0 is presented as a world-model simulator that predicts future observations from visual context and robot actions, with the aim of supporting downstream learning and decision-making. Its technical report describes synchronized numerical and visual representations of actions. 17
That makes its stated purpose different from Pelican-VLA 0.5’s: rather than directly describing a model that maps perception to action, Pelican-Sim focuses on predicting future observations conditioned on actions. The report supports that model role, but the provided evidence does not establish the keynote’s claimed four-layer data strategy or the detailed success-rate results for the proposed 50-plus-500 trajectory setup. 17
Pelican-Unify: combine embodied understanding and generation
Pelican-Unify 1.0 is described in its technical report as a unified embodied intelligence model for understanding, reasoning, imagination and action. The report says a single vision-language model serves as a unified understanding component, mapping scenes, instructions, visual context and action histories into a shared semantic space. 18
The model was already documented publicly before the October event: its report carries a May 2026 identifier, and a May 2026 news report describes its release. 18
24 The evidence therefore does not support calling Pelican-Unify newly introduced in October in the sense of being first made public then. It does not rule out the possibility that it was newly discussed with that conference audience.
How the models fit together
Taken together, the reports suggest complementary pieces of an embodied-learning toolkit: Pelican-VLA 0.5 links visual-language inputs with predicted action and future frames; Pelican-Sim 1.0 models future observations under robot actions; and Pelican-Unify aims to unify understanding, reasoning, imagination and action. 17
18
44 That is a synthesis of the models’ published descriptions, not a verified sequence or roadmap announced at the keynote.
What remains unverified about the keynote
The ROSCon China 2026 website confirms that the conference was scheduled for October 16–18 in Shanghai. The available conference listing does not establish the keynote’s title, speaker, October 17 slot or content. 3
The available sources also do not confirm the keynote’s exact definitions of the four proposed generalization levels—background, compositional, cross and out-of-domain—or the claimed progression from robot-body-free first-person video and handheld UMI capture through simulated data, failure-recovery data and feedback-driven policy iteration. Without the talk recording, slides or a detailed official recap, assigning those definitions and steps to Zou Shilong would risk presenting unverified details as fact.
Likewise, the provided evidence does not establish the alleged representation-analysis methods, attention-pathway interventions and behavioral tests for isolating visual and language reliance. Nor does it substantiate the proposed long-horizon failure-recovery work or a specific loop connecting in-model rollouts, policy updates, real-world validation and feedback data.
The best-supported conclusion is narrower: X-Humanoid’s public reports explore ways to connect perception, imagined futures and action, alongside world-model simulation. The more detailed keynote roadmap remains unverified from the available evidence.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "X-Humanoid’s Pelican Models Point Toward Unified Robot Intelligence"?
X Humanoid’s Pelican reports show complementary approaches to robot action, world simulation and embodied reasoning.
What are the key points to validate first?
X Humanoid’s Pelican reports show complementary approaches to robot action, world simulation and embodied reasoning. Pelican VLA 0.5’s report integrates vision language understanding, future frame generation and action prediction; it reports 91.4% success on RoboTwin Clean and 91.0% on RoboTwin Randomized after fine tuning.
Sources
arxiv.orgPelican-Sim 1.0: A General World Model Simulator for Embodied ...
github.comreadme.md - Pelican-VLA 0.5
huggingface.coREADME.md · X-Humanoid/Pelican-VLA05 at 5623cf1629aae870b973025a675e4534adc44d83