| license: mit | |
| tags: | |
| - robotics | |
| - video-prediction | |
| - recap | |
| # ReCAP robot reference checkpoints | |
| Matched tokenizer, world model and action ranges for RT-1, CALVIN D-to-D, LIBERO-90 and BridgeData V2. Eight exports, 36 asset files, approximately 4.14 GB. World models have 127,101,696 parameters; tokenizers have 129,448,016 parameters. All exports were SHA-256 verified and strictly loaded on CPU on 2026-09-09. | |
| ReCAP itself is training-free and adds no weights. Use each dataset's paired models and action ranges. These checkpoints predict robot video with supplied actions; they are not validated deployed controllers. | |
| See the [code repository](https://github.com/Alexander-wu/ReCAP) for installation, dataset links, protocols, results, validation and limitations. The directories under `weights/` have individual cards documenting initialization and adaptation. These exports do not include optimizer or RNG state and are not resumable training snapshots. | |
| The MIT model terms retain upstream iVideoGPT and RLVR-World attribution. Dataset terms remain with the respective publishers. The ReCAP source code is separately licensed Apache-2.0. | |
| ## Download | |
| Clone [ReCAP](https://github.com/Alexander-wu/ReCAP), install its assets extra, then run `recap download --dataset calvin --root .`. The code release manifest pins an immutable Hub revision and verifies SHA-256 values. Use `rt1`, `calvin`, `libero`, or `bridge` to select a matched pair. | |
| ## Version | |
| The verified model tensors are pinned to `61f3c350349f58ba8619c2454b5fac5802756282` in the release manifest. Later model-card updates do not alter that immutable asset revision. | |