| license: other | |
| tags: | |
| - robotics | |
| - vision-language-action | |
| - memory | |
| - imprint | |
| # IMPRINT Models | |
| This repository contains the public model checkpoints for IMPRINT. | |
| ## Layout | |
| | Path | Benchmark | Memory encoder | | |
| | --- | --- | --- | | |
| | `simulation/rmbench/full_attention/` | RMBench | Full attention | | |
| | `simulation/rmbench/recurrent/` | RMBench | Recurrent gated delta | | |
| | `simulation/rmbench/pi0/` | RMBench | Full attention + native Pi0 | | |
| | `simulation/rmbench/molmoact2/` | RMBench | Full attention + MolmoAct2 | | |
| | `simulation/robocasa/full_attention/` | Customized RoboCasa | Full attention | | |
| | `simulation/robocasa/recurrent/` | Customized RoboCasa | Recurrent gated delta | | |
| The Pi0.5 directories preserve the files from their original release | |
| repositories. The MolmoAct2 directory contains a model-only conversion of the | |
| paper checkpoint. Optimizer, dataloader, scheduler, and trainer state are | |
| omitted. See | |
| `release_manifest.json` for immutable provenance, file counts, byte counts, | |
| and checksums. | |
| Code, configuration, and evaluation instructions are published in the | |
| `Amazon-FAR/imprint` GitHub repository. | |
| The checkpoints remain subject to their applicable upstream model licenses | |
| and terms, including Gemma/PaliGemma where used. The repository metadata does | |
| not relicense upstream weights. | |