Introducing RoboJEPA, the first scaling law for multiembodiment world models. We scaled JEPA predictors up to 8B params, showed scaling laws extrapolate and correlated compute scaling with planning performance. We release all checkpoints and the code.

RoboJEPA is a deterministic (non-diffusion!) world model that predicts the future in the V-JEPA 2.1 latent space, given actions. Uses two losses: 1. predict the next latent from real past latents+actions. 2. predict the next latent from self-predicted past latents+actions.

We establish scaling laws for world-model imagination error as a function of training compute and verify that they extrapolate. Interestingly, we discovered a new type of scaling law: L(C) = E + A·C^(α − γ·ln C)

We train a decoder on top of V-JEPA 2.1 latents and use it to decode action-conditioned imagination into pixels. This visualizes what scaling actually improves: with more parameters and compute, we get consistent robot bodies, better object interaction, and longer predictions.

Our main goal is to show that a scalable world model can plan zero-shot on a real robot toward a single goal image. We plan with the cross-entropy method over a multi-step horizon. There is no policy: actions are sampled from a uniform proposal.

First, we study how planning performance improves with training compute on four simple RoboCasa tasks. We find that planning performance scales predictably, and the world models trace a compute-optimal frontier in planning success.

Next, we deploy the same planner on a real-world Franka robot and test it on three tasks: grasping, object lifting, and simple pick-and-place. With the same planning hyperparameters for all tasks and model sizes, we observe that success improves with model size on all three tasks.

We release all 72 model checkpoints, trained at different sizes and with different amounts of compute. We also release the code for training, evaluation, and planning. Website: Code: paper:

Thanks to my collaborators and advisors. This work would not have been possible without them: @nicolasvelezb , @DYDYYDYYYD, @apsarathchandar , @TusharNagarajan, @_dsevero, @koustuvsinha, Michal Drozdzal, @adri_romsor, @leto__jean , Nicolas Ballas, Mido Assran!

Full 5 minutes video:

Same but vertical

Great work, Artem! I’m Junhan, a seasoned JEPAs. We’ll release a special, solid JEPA representation learning work in mid-to-late Oct. I believe it can significantly accelerate and optimize training for large and future ultra-large models. Stay tuned—hope to connect and discuss!

Looking forward!

@artemZholus RoboJEPA: primera ley de escala para world models multi-cuerpo, predictores JEPA hasta 8B sin difusión. Si el cómputo de planificación crece de verdad con el tamaño, eso sirve más a un robot real que otra demo de feria. Bien por soltar checkpoints y código.

Proving that offline latent imagination error directly correlates with real-world robot success is massive—it finally gives robotics the predictable scaling feedback loop LLMs had.

correlating compute with planning performance is the useful bit. world model papers usually stop at prediction loss

@ylecun October is the month of JEPA!

Scaling JEPA predictors to show compute scaling tracks planning performance is compelling, but the open question is whether that correlation holds when planning has to recover from prediction errors compounding over long horizons.

We actually address that in one of sections of the paper - it does! Please take a look at the paper.

JEPA world models could be a major asset for exploring Titan, Venus and beyond: anticipating the consequences of actions while accounting for both the environment and the spacecraft’s own condition. @artemZholus @ylecun Are you considering space exploration applications for RoboJEPA, or collaborating on a world-model module for robotic explorers like Curiosity or Perseverance?

Congrats on shipping this, robotics feels like it's finally getting its own world model moment, curious how well it holds up on a robot it hasn't seen

why did you stop at 8b though, genuine question

This is my intership paper. We could train 16B but I just did not have enough time of my internship to do a training run that is this long. Also, if you look at the paper, we show that at 8B we are close to saturation already and we need more data.

thank you for your answer

Significance: excellent

Thanks!

@ylecun Big milestone. Finally solid scaling laws for robot world‑models, open‑sourced as well.

Releasing all 72 checkpoints gives other teams a way to test how much model they actually need for their own tasks. That’s a useful thing to put in people’s hands.

Great work!

Thanks!

@ylecun Congrats!

We are building on this as well @OntosWorld but for a unified policy for any type of robots.

This makes wonder if latent future-state prediction could move from detecting risk to anticipating how motion evolves

Exciting! Why the non-commercial license?

We train on open datasets released by several institutions and companies. The datasets released by companies most often come with non-commercial licenses which automatically require all models trained on this data to have the same license. For example AgiBot-beta is like that.

That makes sense for the weights. What about the code?

How can a deterministic world model model a stochastic environment?

By predicting the expected (in probabilistic sense) outcome instead of sampling from the distribution of all possible future outcomes. The stochastic formulation is useful but the deterministic one works as well including the stochastic environments.

I agree that this is the only thing a regression loss can learn, but struggle to understand why the mean future state would a valid state. Have you tried it in non deterministic environments like driving or locomotion?

JEPAの予測器を大きくするだけでなく、計算量と計画性能の関係まで示したのが興味深いです。スケーリング則が外挿するなら、次の規模でどこまで計画能力が伸びるのかも気になります。

The authors say the model is deterministic rather than diffusion-based, predicting the future in V-JEPA 2.1's latent space given actions. With every checkpoint and the code released, outside labs can test whether the scaling curve holds on their own robots.

I fear we might see the same lesson from the LLM literature, where better data and harnesses were actually super important and can be combined with small models.. i guess in this case reducing imagination error, but seems like exciting area to watch

Multi-embodiment scaling up to 8B is the part most papers skip.

I am working on skeleton-based fall anticipation, identifying pre-impact motion before a fall and found 39% detection at 16% FAR with 1.16s mean lead time across unseen rooms.

Learning to “imagine” future states in latent space and then using that for planning feels like a pretty natural direction for embodied AI.

releasing every checkpoint is what makes a scaling law checkable, the number that matters next is whether the planning gains hold on embodiments the model never saw, since that's the case a factory with mixed hardware actually has

Could you please help me understand why your method’s metrics on pick and place are significantly lower than those of pi_0.5?

Love the single goal image. You can show the robot where you want things to end up and leave the intermediate actions to the planner

@ylecun 8B parameter world models hitting scaling laws means embodied AI might follow the same compute-driven trajectory language models took. curious if the data efficiency gap closes too or if physical interaction stays sample-hungry.

First multi-embodiment JEPA scaling law with open checkpoints is the kind of result robotics actually needs — not another closed demo reel. Correlating compute scaling with planning performance is the useful chart. At 8B, what broke first when you moved across embodiments — representation, or the planner?