Orca: The World is in Your Mind
arXiv · HuggingFace · ▲230
Abstract (verbatim)
We introduce Orca, an initial instantiation of a general world foundation model. Orca learns a unified world latent space from multimodal world signals and exposes it through multimodal readout interfaces. Rather than optimizing isolated next-token, next-frame, or next-action prediction, we are centered on Next-State-Prediction modeling, offering a unified state-transition modeling route toward understanding, predicting, and acting upon the world. Orca learns through two complementary paradigms: unconscious learning captures dense natural state transitions from continuous videos, and conscious learning models sparse meaningful state transitions by language-described events and VQA supervision. For pre-training, we construct a large-scale world-learning inventory data, including 125K hours of video data and 160M event annotations. After pre-training, Orca learns a unified world latent space. To examine whether the learned latent supports downstream, we evaluate it by three representative downstream readouts: text generation, image prediction, and embodied action generation. Orca's backbone is frozen, and only the lightweight modality-specific decoders are trainable. Experiments show the scalability of the proposed paradigm and verify that stronger world latent enables stronger downstream readouts. Orca outperforms similar-sized specialized baselines. These results show that Orca, as a general world foundation model, presents a promising approach to understanding, predicting, and acting upon the world. Finally, we discuss the current limitations, aiming to provide useful insights and inspiration for the community.
Background
Background Analysis
1. Technical Context and Real-world Needs
Current AI applications are evolving from single-task systems (e.g., text generation or image recognition) to complex scenarios requiring interaction with the real world. For instance, robots need to combine visual observations and language instructions to perform actions, autonomous vehicles must integrate sensor data and environmental descriptions for decision-making, and virtual assistants require understanding user behavior and semantics for personalization. The core need here is to enable AI to understand the world like humans do—learning from multimodal signals (vision, language, actions, etc.) to predict future states and plan actions. However, existing methods often focus on specific tasks (e.g., next-token or next-frame prediction) without a unified model of the world.
2. Previous Limitations
Traditional AI models (e.g., Next-Token-Prediction or Next-Frame-Prediction) suffer from:
- Fragmented Learning: They optimize for single modalities or tasks (e.g., only text or images), failing to capture dependencies between different signals (e.g., how actions in a video relate to language instructions).
- Lack of Dynamic Modeling: Most models rely on static or labeled data, struggling to learn natural state changes from continuous, unlabeled real-world data (e.g., long videos).
- Task-Oriented Rather Than World-Oriented: Their goal is to complete tasks (e.g., generating images) rather than understanding the world itself, limiting their ability to generalize to unseen scenarios.
These issues restrict performance in complex environments, such as robots failing in unexpected situations or assistants misunderstanding implicit user intent.
3. Orca’s Solution
Orca addresses these problems with a "world latent space" concept:
- Unified Multimodal Learning: It learns a shared latent representation from multimodal data (video, images, language) that encodes the world’s states (e.g., object positions, action consequences).
- Two Complementary Learning Paradigms:
- Unsupervised Learning: Learns "natural state transitions" from continuous videos (e.g., predicting the next frame in a sequence of a person opening a door and picking up keys).
- Supervised Learning: Learns "meaningful state transitions" from language-described sparse events (e.g., "robot picks up a cup"), linking language logic to visual observations.
- Downstream Task Validation: By freezing the main model and training only lightweight decoders (e.g., text generation, image prediction), it验证s the latent space’s generality. Experiments show that a stronger latent space improves downstream performance.
4. Key Differences from Prior Work
Orca’s innovations include:
- From Task-Oriented to World-Oriented: Unlike models optimized for single tasks (e.g., GPT-54 for text generation), Orca aims to model the world’s underlying rules, supporting multiple downstream tasks.
- Dual Paradigm Integration: Combines unsupervised natural dynamics learning with supervised language-conditioned learning to capture both real-world continuity and human logic.
- Scalability Validation: Large-scale data (125K hours of video + 160M event annotations) proves performance improves with data/model size, rather than task-specific optimization.
In summary, Orca represents a step toward general world foundation models, using unified state-transition modeling to truly understand and interact with complex environments.



