Latest in AI
RynnUnify Releases Visual Planning and Agentic RL Code: What Developers Can Use
Alibaba DAMO Academy's embodied AI project now has public implementation code, with a useful distinction between imagining a future and knowing when that helps.
Alibaba DAMO Academy's RynnUnify project released visual-planning and agentic reinforcement-learning code on October 25. The release gives researchers implementation paths for a model that combines reasoning with video generation, including a way to train it to decide when imagining a scene helps answer a question.
The timing matters here. The paper appeared on October 23, and the checkpoint repositories were updated on October 24. The new development is the October 25 code release, rather than a new model announcement today.
What the RynnUnify code release includes
The public repository has two main implementation areas. VisualPlan provides an interleaved visual-planning workflow. BCPO provides training and evaluation code for Branch-Conditional Policy Optimization, the project's approach to teaching a reasoning agent when to use video generation.
The repository also links three public checkpoints: RynnUnify-Base, RynnUnify-Agentic-Reasoner and RynnUnify-WAM-RoboCasa. Their Hugging Face repositories contain downloadable weight files, rather than only model descriptions. These are research components to assemble into a workflow, not a hosted product with a documented per-token API price.
The base model combines an 8-billion-parameter autoregressive reasoner with an 8-billion-parameter rectified-flow video generator. The reasoner predicts the next subtask and describes what should happen. The generator predicts the corresponding visual sequence, using reasoning features alongside scene and motion context.
For a planning rollout, the system alternates between text instructions and generated video. The last predicted frame becomes the observation used for the next step. That gives the model a way to carry a possible future forward, but also means an incorrect prediction can influence later steps.
Why RynnUnify learns when to generate video
Generating a video for every question would add work even when the original observation already contains the answer. RynnUnify's agentic training explicitly separates direct reasoning from reasoning assisted by generated video.
The default training setup samples four direct trajectories and four generation-assisted trajectories for each routable question. It compares rewards within each branch to improve answers, then uses the difference between branches to train the routing decision. The generator stays frozen while the reasoning agent is updated.
The useful distinction is whether generation improves the answer, not whether the model can generate something plausible. For a question about what happens after an object moves, imagined frames may provide useful evidence. For a question answered directly by the current image, generating a future could be unnecessary.
This also makes evaluation more demanding. A working system needs to be assessed on answer quality, routing decisions and the extra computation required by the assisted path. A better combined score does not establish that every request benefits from generation.
RynnUnify benchmarks and their limits
The project's evaluation tables report the following results. These are the authors' measurements, not an independent replication.
| Evaluation | Reported comparison |
|---|---|
| Rynn-VP visual planning | RynnUnify 54.80, compared with Cosmos 3 Nano 45.05 |
| Rynn-GEU understanding | Direct reasoning 44.0, compared with agentic reasoning 50.0 |
| RoboCasa365 control | RynnUnify-WAM 51.5% success across 20 selected composite tasks |
The visual-planning result is a 9.75-point improvement in that evaluation. The six-point difference between direct and agentic understanding supports the case for assisted reasoning on the tested questions. Neither number establishes a general improvement across arbitrary video or robotics tasks.
The control result has an especially important boundary: it covers selected tasks in a benchmark environment. It should not be read as a success rate for general-purpose robots working in unfamiliar homes.
What developers need before trying it
The documented workflows expect Linux, a CUDA-compatible PyTorch setup, compatible checkpoints and task data. The agentic RL path also needs a frozen generator service, with a judge service for relevant reward tasks. Visual planning requires the compatible Cosmos framework and prepared episode inputs.
The base checkpoint card lists roughly 31.5 GB of weight shards and a custom code dependency. That file size is not a minimum GPU-memory requirement. Runtime memory, intermediate video representations and training state add separate costs.
Licensing also needs a closer read than the repository's top-level badge. Project-authored code and documentation use Apache 2.0, while some VisualPlan sources retain OpenMDW 1.1 headers. The model cards identify a separate Alibaba DAMO Academy license. Publicly downloadable weights do not make every component Apache-licensed.
For research teams, this release moves RynnUnify from a set of results toward an inspectable training and planning workflow. The next useful test is whether its routing policy can reserve expensive visual imagination for the questions where that extra computation earns its place.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn research, launches and product decisions into clear stories that developers can understand and share.