# BWM turns Wan2.2-TI2V-5B into an action-conditioned simulator, with training code still unreleased

> Boundless-World-Model is an action-conditioned video world model for robotic manipulation, built on a 5B video backbone and evaluated on the WorldArena challenge. Inference code, model definition and weights are out; training code is not, which caps what you can do with it.

**boundless-large-model/boundless-world-model** — High-fidelity world models for general embodied intelligence, such as data engines and world  simulators.

- Repository: https://github.com/boundless-large-model/boundless-world-model
- Stars: 1,838 · Forks: 76
- Language: Python
- License: Apache-2.0
- Published: 2026-10-16 · Updated: 2026-10-16 · Language: en
- Canonical page: https://hysenlabs.com/projects/boundless-large-model-boundless-world-model

## What the model does: predict the next chunk of frames given actions

BWM is described as a physically consistent, action-conditioned video world model built on Wan2.2-TI2V-5B, positioned as a low-cost simulator for robotic manipulation. The intended use is generating video that a policy could have been trained or evaluated against, rather than controlling a robot directly.

The mechanism is autoregressive chunk prediction. Starting from an initial observation, a dynamic history and an action chunk, the model predicts future observation chunks. The key architectural choice is how actions enter: robot actions are fed into the video diffusion backbone as cross-attention tokens alongside action-conditioned timestep embeddings.

That arrangement is what makes the model action-conditioned rather than a generic video generator. Action conditioning is what distinguishes this setup from generating video without action inputs. Cross-attention on the action chunk ties the generated frames to what the robot was told to do.

The framework covers three areas, in the README's ordering: action-aligned data construction, action-conditioned world simulation, and downstream robot-learning applications.

## Environment assembly: pinned torch, pinned diffsynth, two dependency lists

The requirements file gives explicit install instructions in comments, which is unusual and helpful, since the ordering matters:

```bash
#   pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
#   pip install diffsynth==2.0.11
#   pip install -r requirements.txt
```

The recommended environment name is BWM. The torch install is pinned to 2.8.0 with the cu128 index, so this is a CUDA 12.8 setup and not something that will run as documented on a CPU-only machine.

The remaining inference dependencies are pinned exactly: numpy==1.26.4, omegaconf==2.3.0, imageio==2.37.3, imageio-ffmpeg==0.6.0, pyarrow==23.0.1, Pillow==12.1.1, einops==0.8.2, tqdm==4.67.3 and safetensors==0.7.0.

Note the split between pyproject.toml and requirements.txt. The package metadata declares the distribution as wan-video-action version 0.1.0, describes it as action-conditioned WAN video generation using DiffSynth-Studio, requires Python 3.10 or later and depends on diffsynth>=2.0.0, with the comment that DiffSynth-Studio provides all base dependencies including torch and einops. The requirements file instead pins diffsynth==2.0.11 and leaves torch out of the list entirely.

The package metadata and requirements file specify different DiffSynth versions, so they do not describe exactly the same environment. They will not necessarily produce the same setup, and the README does not give a full inference command, so treat the demo directory and docs as the place to look next.

## WorldArena results, and what the leaderboard categories mean

The evaluation is the CVPR 2026 WorldArena challenge, hosted as a leaderboard on Hugging Face Spaces. Two entries are reported: BLM, which took first place among open-source models on Track 1 and on Track 2 Data Engine, and BWM-fast, which took second place on the overall Track 1 leaderboard.

The distinction between those two claim types is worth keeping straight. First among open-source models on a track is not the same as first overall, and the leaderboard images are separated into Track 1 open-source, Track 2 Data Engine open-source and Track 1 overall for that reason.

The generated scenes are described as produced autoregressively from initial frames and action sequences in the WorldArena test set. The qualitative breakdown groups them into five scenario families: compositional spatial rearrangement such as arranging blocks by size and stacking bowls, articulated hinge interaction such as opening a microwave or laptop, fine-grained affordance interaction such as turning a switch, hanging a mug, clicking a bell and stamping a seal, bimanual coordination and handover, and long-horizon constrained placement such as putting objects in a cabinet or bottles in a dustbin.

Each family names a failure mode the model is claimed to avoid: stacking stability, hinge-constrained rotation, contact regions, inter-arm occlusion and object drift under partial occlusion. Those are claims about generated video, not measurements of physical accuracy, and the README does not report a physical consistency metric.

## Release state: everything except training code

The README carries a checklist, and one box is empty. Inference code is released, the model definition is released, the model weights are released, and the technical report is released. Release training code is listed as not done.

That single gap defines the audience. You can run the model and you can read the architecture. You cannot reproduce the training run from this repository, and with no GitHub releases tagged there is no versioned artefact trail beyond the model card on Hugging Face under BLM-Lab.

The news timeline gives a sense of the order things appeared: the model definition and inference code in 2026-05 alongside the leaderboard results, and a technical report in 2026-07. The last push was on 2026-09-09 and the repository is not archived.

Licensing is Apache-2.0 in pyproject.toml, and the repository carries a standard LICENSE file.

## Repository layout suggests where the data pipeline lives

The top level is small: assets, configs, demo, docs, scripts, the wan_video_action package, plus pyproject.toml and requirements.txt.

Several entries point at the data side of the framework. demo holds a demo.jsonl and a stat.json alongside per-scene directories such as demo/adjust_bottle, which is consistent with a JSON Lines set of generation requests plus per-demo outputs. pyarrow appears in the pinned dependencies, and it is a columnar format library, which suggests structured record handling somewhere in the generation or dataset path.

configs and scripts sit alongside, so configuration for the generation runs is separated from the library code. The README does not explain the schema of demo.jsonl or the contents of stat.json, so anyone writing automation against them needs to read the files directly rather than guess.

The framework figure described as overview of BWM framework is in the README as an image, so the data construction and downstream robot learning stages are described in text only at the level of the three-part summary.

## Where it is the wrong tool

This is a video world model for manipulation scenes, not a robot control stack and not a general simulator. It produces pixels conditioned on actions; it does not output state estimates, contacts or torques, and nothing in the README suggests otherwise. A team looking for a physics engine to plan against is looking at the wrong repository.

The cost profile is the other practical limit. A 5B-parameter video diffusion backbone generating chunks autoregressively is a GPU workload, with a CUDA 12.8 torch build and exact pins that leave little room for substitution. The project describes BWM as a low-cost simulator, but it does not provide comparative cost measurements.

Teams needing a physics engine may need to evaluate a different repository. For evaluating contact-rich stacking or hinge dynamics, the learned route reproduces appearance without reproducing forces.

A second alternative is using the released model as a data engine only, generating rollouts and training a policy on them, and ignoring the physical consistency claim entirely. This downstream use is suggested by the framework summary.

## Conclusion

Adopt Boundless-World-Model if you need action-conditioned video rollouts of manipulation scenes for evaluation or data generation and you already have GPU capacity and a DiffSynth setup. Do not adopt it if you intend to retrain or fine-tune the model, since the README lists training code as the one unchecked item in its release plan. Before committing, verify the checkpoint you pull matches the Wan2.2-TI2V-5B base and the pinned diffsynth 2.0.11 version the requirements file names, because the package metadata and the requirements file disagree on how the environment is meant to be assembled.

## FAQ

### Is the training code for Boundless-World-Model released?

No. The README checklist marks inference code, model definition, model weights and the technical report as released, while release training code remains unchecked.

### What base model does BWM build on?

BWM is built on Wan2.2-TI2V-5B, a 5B video backbone, and the model is described as an action-conditioned video world model for robotic manipulation.

### How do robot actions enter the BWM model?

Actions enter the video diffusion backbone as cross-attention tokens and as action-conditioned timestep embeddings, and the model autoregressively predicts future observation chunks from the initial observation, dynamic history and action chunk.

### What results does the project report on WorldArena?

In the CVPR 2026 WorldArena challenge, BLM ranked first among open-source models on Track 1 and on Track 2 Data Engine, while BWM-fast ranked second on the overall Track 1 leaderboard.

### What environment does Boundless-World-Model need for inference?

The requirements file recommends installing torch==2.8.0 with torchvision 0.23.0 and torchaudio 2.8.0 from the cu128 index, then diffsynth==2.0.11, then the pinned requirements. The recommended environment name is BWM and Python 3.10 or later is required.

## Sources

- [boundless-large-model/boundless-world-model on GitHub](https://github.com/boundless-large-model/boundless-world-model)
- [Issues](https://github.com/boundless-large-model/boundless-world-model/issues)
- [License: Apache-2.0](https://github.com/boundless-large-model/boundless-world-model/blob/main/LICENSE)
- [README](https://github.com/boundless-large-model/boundless-world-model/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/boundless-large-model-boundless-world-model