pi05_franka_5task_yhy-project_full_10k

π0.5 Dual-Franka Five-Task YHY Project Full SFT — 10k

English | 中文说明

This repository contains the final JAX/Orbax checkpoint from full-parameter supervised fine-tuning of the official OpenPI pi05_base model on five real-robot manipulation tasks collected with a dual Franka FR3 setup.

Tasks and prompts

Use the exact prompt wording above for the closest match to training-time conditioning.

Data

• Platform: dual Franka FR3, joint-space control
• Source: real-robot teleoperation trajectories
• Training split: 223 valid episodes
• Held-out split: 25 episodes
• Per-task training counts: 44 cabinet, 45 plate stacking, 44 bowl-on-plate, 45 cup-on-tray, 45 tape-in-basket
• Held-out count: 5 episodes per task
• Cameras: chest, left wrist, right wrist
• Robot state: 16 dimensions — left 7 joints + left gripper + right 7 joints + right gripper
• Source action: the same 16-dimensional dual-arm joint/gripper ordering
• Source streams were aligned at 20 Hz and stored in LeRobot format

The OpenPI model configuration pads the 16-dimensional source action to action_dim=32 internally and predicts an action horizon of 20.

Training configuration

The last logged training loss at step 9,900 was 0.0050. Training loss is not a real-robot success metric.

Download and loading

hf download THU98/pi05_franka_5task_yhy-project_full_10k \
  --local-dir ./pi05_franka_5task_yhy-project_full_10k

The repository is an OpenPI JAX/Orbax checkpoint. Use it with a matching local OpenPI configuration named pi05_franka_5task_cup_full:

uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi05_franka_5task_cup_full \
  --policy.dir=./pi05_franka_5task_yhy-project_full_10k

The local configuration must match the training architecture (pi05=True, action_dim=32, action_horizon=20) and must supply the included normalization assets. This is not a Transformers from_pretrained() checkpoint.

Evaluation status

No controlled real-robot success-rate evaluation is published with this checkpoint. The held-out episodes were retained for offline checkpoint comparison, but training loss alone does not establish deployment performance.

Intended use and limitations

• Research use on a dual Franka setup with matching cameras, joint order, gripper convention, preprocessing, normalization, control rate, and prompts.
• Not validated for other robots, viewpoints, camera calibrations, objects, or language paraphrases.
• The task set is narrow and consists of left-arm manipulation behaviors; it is not a general-purpose policy.
• Plate/bowl terminology follows the task labels and training prompts shown above.
• Before hardware deployment, verify action bounds, emergency stops, collision avoidance, workspace limits, latency, and low-speed dry runs.
• Human supervision and independent safety controls are required.

中文说明

这是基于官方 OpenPI pi05_base，使用双 Franka FR3 真机遥操数据进行全参数 SFT 得到的五任务模型。模型包含柜顶放碗、叠两个盘子、碗放盘子、杯子放托盘和胶带放篮子。训练使用 223 条有效轨迹，另外保留 25 条轨迹，总训练 10,000 步。

输入包括胸部、左腕、右腕三路图像、Language Instruction 和 16 维双臂关节/夹爪状态；输出为双臂动作序列。当前没有发布正式真机成功率，部署前必须核对相机、动作顺序、控制频率和安全限制。

Sources

• OpenPI: https://github.com/Physical-Intelligence/openpi