| --- |
| tags: |
| - futureworlds |
| - robotics |
| - world-model |
| - video-prediction |
| - reinforcement-learning |
| --- |
| |
| # FutureWorlds research checkpoints |
|
|
| FutureWorlds learns action-conditioned robotic world models from alternative predicted futures. This repository contains the paired research checkpoints for RT-1, BridgeV2, and RoboCasa. The implementation, protocols, and reproduction instructions are in the [FutureWorlds GitHub repository](https://github.com/Alexander-wu/FutureWorlds). |
|
|
| ## Contents |
|
|
| Each dataset has a `main` predictor, an SFT predictor, and `memspo200`, `grpo200`, and `ordinary200` comparison predictors. The dataset directory also contains its paired visual codec, action-range metadata, and configuration. RT-1 and RoboCasa predictors include trained text adapters. The original FP32 safetensors are kept without quantization or variant merging. |
|
|
| | Dataset | Main checkpoint | Text conditioning | |
| | --- | --- | --- | |
| | RT-1 | MemSPO, 500 updates, R0 | T5-base plus trained adapters | |
| | BridgeV2 | MemSPO, 200 updates, E4/R1 | None in this recipe | |
| | RoboCasa | MemSPO, 400 updates, R0 | T5-base plus trained adapters | |
|
|
| The external T5-base model is not copied into this repository. Use revision `a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1` of [`google-t5/t5-base`](https://huggingface.co/google-t5/t5-base) for the text-conditioned datasets. |
|
|
| ## Loading |
|
|
| These are custom action-conditioned predictors, not generic `pipeline("text-to-video")` models. Use the dataset-specific codec, action ranges, and text adapter. The included `code/` directory contains the paired low-level checkpoint loader; see the [GitHub loading guide](https://github.com/Alexander-wu/FutureWorlds/blob/main/docs/LOADING.md) for the full RGB pipeline and the [model cards](https://github.com/Alexander-wu/FutureWorlds/tree/main/model_cards) for variant details. `main` and `memspo200` are different checkpoints. |
|
|
| ## Validation and limitations |
|
|
| The exact uploaded-file hashes are recorded in `manifest.json`; `weights_manifest.json` records component provenance and marks whether each listed hash belongs to a source file or a release file. The code repository reports first-frame agreement with the original entry point on one fixed case per dataset. Full-horizon GPU reproduction of the consolidated release pipeline remains pending. Results in the paper and GitHub README are existing experiment results, not a fresh evaluation of this upload. The repository does not contain raw RT-1, BridgeV2, or RoboCasa datasets, optimizer states, or third-party baseline weights. |
|
|
| ## Rights and attribution |
|
|
| No license has been selected for the original FutureWorlds additions or these checkpoints. Public visibility does not grant permission to use, modify, or redistribute them. Upstream components and the external T5 model retain their own licenses; see `licenses/` and [third-party notices](https://github.com/Alexander-wu/FutureWorlds/blob/main/THIRD_PARTY_NOTICES.md). Contact the project owner for permissions. |
|
|
