Latent world models have shown a remarkable ability to predict future states and to plan in the real world. In practice, however, we lack a principled way to estimate how their capabilities scale with model size, data, and compute, an open problem that slows progress in the field. In this work we present RoboJEPA, a world model based on the Joint Embedding Predictive Architecture (JEPA) and trained on a large-scale dataset spanning 12 robotic embodiments. We show that RoboJEPA's imagination error, the error of its latent rollouts, follows a second-order power law in compute, allowing us to predict model quality well beyond the scale at which the law is fit. We further show that downstream robotic planning performance improves predictably with compute, and that imagination error is strongly correlated with it, making it a reliable proxy for real-robot evaluation. Finally, we demonstrate that latent world models can be deployed zero-shot as robotic agents, planning toward a single goal image to solve tasks requiring long-horizon planning on real hardware. We release all model checkpoints together with our training and robot deployment code. To our knowledge, this is the first work to establish scaling laws for multi-embodiment robotic world models trained on real robot data, and RoboJEPA, at 8B parameters, is the largest JEPA predictor model trained to date.
Why scaling predictability matters for robotic world models
Scaling has been one of the primary engines of recent progress in artificial intelligence. What makes scaling particularly actionable is predictability: empirical scaling laws relate a model's loss to the amount of data, parameters, and compute used to train it, so that the return on an additional order of magnitude of compute can be forecast before it is spent. This predictability turns scaling from a gamble into a science, telling practitioners how large a model to train for a given budget, how much data to pair it with, and when returns will diminish. Scaling laws of this kind have been established across language models, generative modeling of images, video, and other modalities, or semantic classification of images. Comparatively little is known about whether, and how, the same scaling predictability holds for models of the physical world.
We study this question for world models in robotics. A world model is a learned model of environment dynamics that an agent can roll out to imagine future states and plan. We focus on latent world models that predict and plan in a latent representation space. Latent world models have shown the ability to plan toward goals specified at test time without the need of a task-specific policy. Previous latent world models are typically trained at a single scale, and how their quality changes with data, parameters, and compute has not been characterized. Other studies, focusing on scaling, either target generative video and interactive-environment models outside of robot control, or scale policies rather than world models. In contrast, our goal is to establish scaling laws for latent world models. This matters particularly in robotics, where high-quality interaction data is scarce and computationally expensive to collect: knowing which capabilities a world model will have at a given training-compute budget lets practitioners decide whether to invest in a larger model, more demonstrations, or more compute, and reveals when a data budget has been saturated.
RoboJEPA: a scalable family of latent world models
To study the scaling behavior of robotic world models, we require a model family that scales with training compute. To this end we introduce RoboJEPA, a family of action-conditioned latent joint-embedding predictive world models. We keep the training recipe deliberately simple to focus on the scaling behavior. RoboJEPA is a feed-forward transformer that predicts the next latent state deterministically in a single forward pass. It is trained with a plain ℓ1 regression loss in the frozen encoder's latent space. RoboJEPA operates in the representation space of a frozen V-JEPA 2.1 encoder and is trained on a unified mixture of 23 public manipulation datasets spanning 12 embodiments and 26 different action spaces. We study model scaling of the world model, ranging parameter count from 22M up to 8B, while holding the representation encoder fixed, so our laws characterize how a plannable dynamics model scales on top of a shared representation.
RoboJEPA models operate in the latent space of a frozen V-JEPA 2.1-G encoder. At each timestep, given a history of past visual features, proprioceptive states, and actions, the model deterministically predicts the features of the next frame in one neural network forward pass. Actions are either state deltas or absolute robot commands. Feeding predictions back as inputs lets us roll the model out autoregressively over future actions, enabling planning.
Scaling laws: imagination error follows a second-order power law in compute
Our central finding is that robotic world-model quality scales predictably with compute. First, the latent forward-prediction loss on out-of-distribution data follows a scaling law L(C) = A C^(α-γ ln C) + E, where C is the training compute, E is the irreducible loss, and A, α, γ are scaling-law parameters that we estimate alongside E. This holds on both real-robot holdout (DROID) data and simulated holdout (RoboCasa) data. Second, this improvement on the offline data transfers to downstream planning performance on a robot: on a suite of RoboCasa planning tasks, downstream success improves predictably with training compute, so that a world model's latent forward-prediction loss is a strong predictor of its online planning performance.
For example, only models trained with more than 10^22 FLOPs of training compute (at sizes 2B, 4B, and 8B) can solve object-interaction tasks in RoboCasa via non-greedy planning, while smaller models trained with less compute cannot. The compute-optimal frontier (lower envelope) follows the second-order power law L(C) = E + A C^(α-γ ln C) (dashed). This second-order term γ ln C captures diminishing returns: each additional order of magnitude of compute yields smaller absolute gains in prediction accuracy.
Downstream planning improves predictably with compute
On a suite of RoboCasa planning tasks, downstream success improves predictably with training compute. Across two robot platforms, we perform planning in over 50,000 evaluation episodes. We keep the planning hyperparameters fixed for each robot, varying only the planning horizon for harder tasks. We perform planning on a set of real world manipulation tasks that include grasping, lifting, and pick-and-place. In simulation, our planning spans tasks of increasing difficulty from reaching to object pushing. We observe an improvement trend in the success rate on the real robot as we increase the model size. Qualitatively, we also observe that the behavior of the planning policies changes as the model grows: for example, in object-manipulation tasks, only models with 2B parameters or more keep the gripper closed after grasping to complete the task, while smaller models often accidentally open it, dropping the object.
Imagination error as a reliable proxy for real-robot evaluation
We further show that downstream robotic planning performance improves predictably with compute, and that imagination error is strongly correlated with it, making it a reliable proxy for real-robot evaluation. World models with lower latent forward-prediction loss consistently achieve higher planning success rates on real hardware. This correlation holds across model sizes and across the two robot platforms evaluated, allowing practitioners to assess world model quality through offline evaluation alone, without the need for costly real-robot deployment.
Zero-shot deployment as robotic agents
We demonstrate that latent world models can be deployed zero-shot as robotic agents, planning toward a single goal image to solve tasks requiring long-horizon planning on real hardware. Across two robot platforms, we perform planning in over 50,000 evaluation episodes. We keep the planning hyperparameters fixed for each robot, varying only the planning horizon for harder tasks. We perform planning on a set of real world manipulation tasks that include grasping, lifting, and pick-and-place. The model plans toward a single goal image that defines the final state of the environment, without any task-specific policy fine-tuning. This zero-shot capability shows that a single world model, trained on diverse multi-embodiment data, can generalize to control new robotic platforms at test time.
Training and model details
We train a series of RoboJEPA models that range in size from 22M to 8B parameters and in training compute from 2×10^19 to 9.5×10^22 FLOPs. For all models, we use AdamW with a weight decay of 0.1 and global gradient-norm clipping at 1.0. We linearly warm up the learning rate, hold it constant at a peak of 5×10^-5 for most of training, and then linearly decay it to zero over a short cooldown phase at the end. This allows us to overlap the flat learning-rate stage across models trained with different amounts of compute: we run the flat stage only once and perform learning-rate annealing from intermediate checkpoints.
The training recipe of RoboJEPA is designed to facilitate scaling the model on as much data as is available, which requires training on many heterogeneous robotic datasets at once. A central issue with such heterogeneous data is that some datasets supply only a single view while others provide multiple views, often at different resolutions. We therefore train the model to work with a variable number of views and a variable resolution. To achieve this, we partition the GPUs into groups, where all GPUs within a group work with the same number of views and the same resolution. In practice this often leads to imbalanced step times across groups, which is why we balance these times by using different batch sizes for different groups.
Related work
Scaling has been one of the primary engines of recent progress in artificial intelligence. Scaling laws of this kind have been established across language models, generative modeling of images, video, and other modalities, or semantic classification of images. Our work is most closely related to recent efforts to establish scaling laws for robotic systems. Other studies, focusing on scaling, either target generative video and interactive-environment models outside of robot control, or scale policies rather than world models. Our work also relates to the broader literature on world models and model-based reinforcement learning, where the focus has typically been on individual algorithms rather than systematic scaling characterization.
Conclusion
We introduce RoboJEPA, a family of action-conditioned latent joint-embedding predictive world models, and establish that robotic world-model quality scales predictably with compute. The latent forward-prediction loss on out-of-distribution data follows a second-order power law in training compute, allowing us to predict model quality well beyond the scale at which the law is fit. Downstream robotic planning performance improves predictably with compute, and imagination error is strongly correlated with it, making it a reliable proxy for real-robot evaluation. Latent world models can be deployed zero-shot as robotic agents, planning toward a single goal image to solve tasks requiring long-horizon planning on real hardware. We release all model checkpoints together with our training and robot-deployment code. To the best of our knowledge, this is the first work to establish compute-optimal scaling laws for action-conditioned robotic world models trained on large-scale, multi-embodiment real robot data, and the first scaling law for JEPA with verified extrapolation. We also train the largest JEPA predictor to date, at 8B parameters.