IMPRINT RoboCasa Recurrent Checkpoint
Canonical Pi0.5 IMPRINT checkpoint for the customized 12-task RoboCasa benchmark using the recurrent gated-delta memory encoder.
- Stage 1: trained 180,000 updates, selected at 160,000, W&B
far-wandb/Robocasa/ag6ytzrm - Stage 2: 150,000 updates, W&B
far-wandb/Robocasa/1ozs9qy8 - memory tokens: 256 x 1024
- policy horizon: 20
- training scan: chunkwise WY
- inference update: incremental recurrent
- Stage 2: memory encoder finetuned; image tower frozen
The Stage-2 model-level auxiliary weight was zero. The archived history YAML
retains a legacy aux_weight: 0.25 preload flag, which caused auxiliary
targets to be read but did not contribute an auxiliary loss.
Use yilin-wu/imprint-robocasa-600 for training and
yilin-wu/imprint-robocasa-environment for the exact patched RoboCasa 1.0.1
simulator source. This release contains the complete Stage-2 Orbax parameter
checkpoint, including the PaliGemma/Gemma backbone. It is a Gemma model
derivative and is distributed subject to the included Gemma Terms of Use.
Large simulator assets are not duplicated.
The matching training store, patched simulator source, and reproducibility code are publicly available under the IMPRINT release.