YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

pick&place_pi0.5_10k

PI0.5 dual-Franka real-world SFT checkpoint at training step 10,000.

This directory contains the artifacts required to reconstruct the trained policy for inference. It deliberately excludes AdamW and distributed-checkpoint state, which are only required to resume training. It also avoids duplicating the frozen base tensors in a second full state dict.

Contents

  • base_model/: pristine PI0.5 pretrained model.
  • checkpoint/actor/model_state_dict/trainable_weights.index.json: index for eight sharded files containing all 789 trainable tensors at global step 10000 (LoRA, Gemma expert, and action/time heads).
  • assets/frankadataset-lerobot/norm_stats.json: inference normalization statistics.
  • config/inference_config.yaml: resolved model configuration, patched to use paths inside this bundle.
  • config/load_paths.yaml: canonical paths for an inference loader.
  • load_model.py: relocatable, exact-key-checking loader for this bundle.
  • source_snapshot/: the custom dual-Franka policy/data configuration used by this checkpoint, retained for provenance.

Policy schema

  • Task: place the banana and the tape into the basket using both arms.
  • Cameras, in order: cam_high, cam_left_wrist, cam_right_wrist.
  • State/action dimension: 16.
  • Layout: left 7 joints, left gripper, right 7 joints, right gripper.
  • Joint actions are relative; gripper commands are absolute continuous values in [0, 1].
  • Action horizon: 50.

Loading contract

Use the RLinf environment at /data2/zhoujingjing/h100_migration/workspace/RLinf. Construct the model from config/inference_config.yaml, load the eight shards referenced by checkpoint/actor/model_state_dict/trainable_weights.index.json, move the model to CUDA, and switch it to eval mode. The loader requires an exact match with the model's 789 trainable parameter keys. The model builder first reconstructs all 604 frozen tensors from base_model/ and loads the normalization statistics from assets/.

The compact representation is numerically equivalent to the original RLinf full_weights.pt: all 604 omitted frozen tensors (6,902,063,840 bytes) were verified byte-for-byte against the pristine base before publication.

The source snapshot is documentation; the active RLinf checkout must contain the registered pi05_dual_franka data configuration.

Smoke-load the complete bundle with:

/data2/zhoujingjing/h100_migration/workspace/RLinf/.venv/bin/python \
  load_model.py --device cuda
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support