Beijing Academy of AI's RoboBrain Orca bets the next AI paradigm is world-state prediction

BAAI, the Beijing Academy of Artificial Intelligence, has released RoboBrain Orca, a multimodal world model whose loss function does not ask the network to guess the next word, the next frame, or the next robot command. It asks the network to guess the next state of the world.

That collapse of three prediction targets into one is the actual news. Inside the AI industry, next-token prediction has been the default objective since large language models took over. Robotics added next-action prediction for control. Video models added next-frame prediction for synthesis. Orca, described in the lab's preprint and BAAI's conference materials, folds all three into a single objective: predict how the world changes next.

A world state, in this framing, is not a sentence or a frame. It is an abstract internal representation of a scene at a moment in time, the kind of structured snapshot a human carries when remembering "the kitchen at 8pm" without replaying every pixel or word.

To build that representation, BAAI says Orca was trained on roughly 125,000 hours of video and 160 million event annotations, figures the lab cites from its own scaling runs. The training regime splits in two.

The first half is what the lab calls unconscious learning: the model watches raw video of the physical world and learns how objects, people, and scenes move on their own. There are no captions, no instructions, no labels. The goal is for the network to internalize the world's physics and rhythms the way an infant does, by sustained exposure.

The second half is conscious learning. The same model is shown natural-language descriptions of events, task instructions, and visual question-answer pairs. The labels teach it to verbalize what it has already internalized: to answer "what happens next if I drop this?" in language, not just simulate the drop internally.

What changes at inference is the readout head.

A text head produces language describing state transitions and dynamic motion. An image head produces visual predictions of how a real scene will look one moment later. An action head produces robot commands, even though no action labels were used during pretraining. That last head is the technical surprise: BAAI claims the action readout emerges from world-state representations the model built without ever seeing an action label, and transfers to downstream robot policies without re-training on robot data.

That transfer claim is the crux. If it holds, embodied-AI builders could skip the expensive step of collecting action-labeled robot trajectories and lean on a general world model instead. If it does not hold, Orca is another large video model with a robot head bolted on for the press release.

The lab's earlier RoboBrain, published in early 2025, was a reasoning model for embodied AI. RoboBrain 2.0, released later that year, added richer perception and tool use. Calling Orca a world model rather than a robot policy is the sharpest break from the earlier line. Whether Orca supersedes the earlier two or runs alongside them is not spelled out in the public materials.

The wider context is timing. As returns on next-token scaling have visibly flattened across the industry, world models have moved from curiosity to strategic bet for several labs. Orca is the most explicit public attempt so far to rebase an entire research program on a non-token objective. BAAI's 2026 conference program treats world models as a headline topic, suggesting the lab intends to argue next-state prediction as the next scaling axis rather than an Orca-specific trick.

There are reasons to read carefully. The 125,000-hour and 160-million-event figures come from secondary Chinese coverage of the lab's announcement and have not yet been independently verified against the technical report. The unconscious and conscious split is described in narrative terms; the precise data ratios, loss weighting, and benchmark scores for each path are not yet public. And the action head's downstream performance rests on evaluations BAAI controls.

What to watch: an updated arxiv version with full benchmark tables, third-party replication of the action-head transfer, and whether other world-model labs, including Google's DeepMind, Meta, and a handful of well-funded Chinese peers, respond with their own next-state architectures or stay committed to token-side scaling. BAAI is betting the AI industry's next decade will look less like predicting language and more like predicting reality itself.