OxyGen pi0.5 LIBERO textual-memory suffix LoRA
This rank-16 LoRA produces observation-grounded textual memory from the current
observation, robot state, and task instruction. It is enabled for the private
Memory: seed and the autoregressive language suffix. The frozen base model
encodes the shared prefix, and the action expert reads the unchanged root KV
cache.
| Item | Value |
|---|---|
| Base checkpoint | gs://openpi-assets/checkpoints/pi05_libero |
| Suffix seed | Memory: |
| Maximum suffix length | 28 tokens |
| Held-out token accuracy | 98.15% |
| Greedy exact match | 87.25% |
| Greedy word F1 | 90.01% |
| Root-cache/action maximum absolute difference | 0 / 0 |
| Adapter overhead | 8.16 ms |
| Shared prefix forward | 47.66 ms |
Training, evaluation, and serving instructions are in OxyGen under
language_adaptation/libero.
The matching derived annotations are available at https://huggingface.co/datasets/xxxxyu/libero-textual-memory-annotations.
OxyGen/openpi code is Apache 2.0. Gemma components remain subject to the Gemma terms distributed with openpi. The base checkpoint is not included here.