Instructions to use Shiki42/a2r-e8-insert-key-pi05-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use Shiki42/a2r-e8-insert-key-pi05-lora with LeRobot:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
a2r E-8 insert_key π0.5 LoRA (E8-R001)
LeRobot 0.6.1 pi05 with a peft LoRA adapter (r=32, alpha=32) on lerobot/pi05_base, fine-tuned for one contact Skill, key_insert_turn (insert a held key into a lock and turn it), on a simulated dual ARX-X5 (RoboDojo, 25 Hz).
- Data: 200 admitted single-arm expert demonstrations (180 train / 20 validation by layout), dataset
Shiki42/a2r-e8-insert-key@6ce7242012a181088bb6d2862fcfca2ca8b89c5f(private). - Training: 10,000 steps, batch 16, seed 20261018, bf16, gradient checkpointing; final train loss 0.100, fixed-noise validation loss 0.091.
- Observation:
observation.images.top(front) and the active arm's wrist camera, state = active-arm link6 pose in its base (xyz + quaternion xyzw) and two finger widths, the Skill instruction. - Action (7): body-frame EE delta relative to the measured link6 at the executed frame (xyz m, rotation vector rad) and absolute gripper open fraction. Not certified for hardware.
pretrained_model/loads with LeRobot 0.6.1 (use_peft, camerarename_mapintrain_config.json).
Simulation-only research artifact from the a2r project (Experiment E-8).