pi05_songling_plate_stack_v1_lora_s1

LoRA adapter for lerobot/pi05_base, fine-tuned on YSanYi/songling_aloha_plate_stack_v1 for the SongLing ALOHA plate stacking task.

This repository contains the PEFT adapter weights, the policy config, and the pre/post-processing artifacts needed by LeRobot.

What is in this repo

• adapter_model.safetensors: LoRA adapter weights
• adapter_config.json: PEFT adapter config
• config.json: policy config for pi05
• policy_preprocessor*.safetensors: observation/action normalization artifacts
• policy_postprocessor*.safetensors: action de-normalization artifacts
• train_config.json: final training config

Base model

• Base policy: lerobot/pi05_base
• Adapter type: LoRA
• Rank: r=32

Training data

• Dataset: YSanYi/songling_aloha_plate_stack_v1
• Task: plate stacking with SongLing ALOHA

Training setup

• Policy: pi05
• Steps: 15000
• Batch size: 64
• Dtype: bfloat16
• Learning rate: 1e-5
• Warmup steps: 500
• Decay steps: 6000
• Final decay LR: 2e-6
• Gradient checkpointing: true
• Seed: 1000
• Normalization:
  • ACTION=QUANTILES
  • STATE=QUANTILES
  • VISUAL=IDENTITY

Training completed successfully at step 15000.

How to use with LeRobot

Load this adapter on top of lerobot/pi05_base with LeRobot's pretrained policy loading path.

from lerobot.configs.policies import PreTrainedConfig
from lerobot.policies.factory import make_policy

cfg = PreTrainedConfig.from_pretrained("YSanYi/pi05_songling_plate_stack_v1_lora_s1")
policy = make_policy(cfg, pretrained_policy_name_or_path="YSanYi/pi05_songling_plate_stack_v1_lora_s1")
policy.eval()

If your local LeRobot version expects a direct pretrained path, point it at this repo id as the policy checkpoint path.

Notes

• This is a LoRA adapter repo, not a merged full-weight repo.
• You need the peft dependency available in your environment to load it.
• The adapter was trained from the pi05_base initialization and should be used with the same base architecture.

Limitations

• This model card reports training configuration and artifacts only.
• No online eval metrics or rollout success metrics are included in this repo.
• Final model quality should be judged with task-specific rollout evaluation.