pi05-pp9robot-bs128-step7999

A pi0.5 LoRA fine-tune on a 9-embodiment pick-and-place dataset, checkpoint step 7999 of a 20,000-step run, with the LoRA adapters merged into the base weights (no PEFT/adapter loading required).

Loading

This is a raw openpi/JAX (orbax) checkpoint, not a transformers-native format. To load it you need an openpi checkout with a TrainConfig whose model uses the non-LoRA variants (gemma_2b / gemma_300m) — the merged weights have the same shapes as a full fine-tune, so they load like one:

from openpi.training import config as _config
from openpi.policies import policy_config as _policy_config

train_config = _config.get_config("<your_non_lora_config_name>")  # model=gemma_2b/gemma_300m
policy = _policy_config.create_trained_policy(train_config, "<path_to_this_repo>")

assets/ contains the training-time normalization statistics (norm_stats.json), required for correct action de-normalization at inference time.

Caveats

• Trained with no embodiment conditioning — expect uneven performance across the 9 embodiments (that unevenness is one of the things this line of experiments is studying).
• Mid-training snapshot; loss and held-out SR were still moving at this step.
• Data volume per object category varies widely (4 to 108 training episodes); rare objects should be expected to perform worse, independent of embodiment.