IMPRINT RMBench Full-Attention Checkpoint
Canonical Pi0.5 IMPRINT checkpoint for RMBench using the full-attention memory encoder.
- Stage 1: 160,000 updates, W&B
ylinwu/Robocasa/j8t4ebpo - Stage 2: 100,000 updates, W&B
liqhuiha/Robocasa/grruxd5p - memory tokens: 256 x 1024
- policy horizon: 50
- visual memory: three-view pooled DINOv3 features
- Stage 2: memory encoder finetuned; image tower and language embedding table frozen
The release contains the Stage-1 encoder, the complete Stage-2 Orbax parameter checkpoint (including the PaliGemma/Gemma backbone), normalization statistics, portable history configuration, and provenance. The checkpoint is a Gemma model derivative and is distributed subject to the included Gemma Terms of Use.
Use the matching yilin-wu/imprint-rmbench-450 dataset and the evaluation
contract in the IMPRINT GitHub repository. Evaluation requires strict causal
actions, online recent DINOv3 features, EVAL_CONSOLIDATE_ALL=1, BGR policy
inputs, and 50-step prediction/execution chunks.
The matching training store and reproducibility code are publicly available under the IMPRINT release.