π0.5 Dual-Franka Five-Task YHY Project Full SFT — 10k
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
| Task ID | Training prompt |
|---|
place_bowl_on_cabinet | Place the green bowl on top of the cabinet with the left arm. |
stack_bowls_two | Use the left arm to stack the blue plate on top of the green plate. |
place_bowl_on_plate | Place the gray bowl on the green plate with the left arm. |
place_cup_on_tray | Place the cup on the tray with the left arm. |
tape_on_basket | Place the tape in the green basket with the left arm. |
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
| Setting | Value |
|---|
| Base checkpoint | Official OpenPI pi05_base |
| Method | Full-parameter SFT |
| Training steps | 10,000 |
| Hardware | 4 × NVIDIA A100 80GB |
| Global batch | 16 |
| Optimizer | AdamW, gradient clipping at 1.0 |
| Learning rate | cosine schedule, 1,000 warmup steps, peak 1e-5, floor 1e-6 |
| EMA | Disabled |
| Checkpoint interval | 2,000 steps plus final checkpoint |
| Loss logging | Every 100 steps |
| Final checkpoint index | 9999 |
The last logged training loss at step 9,900 was 0.0050. Training loss is not a real-robot success metric.
Download and loading
1hf download THU98/pi05_franka_5task_yhy-project_full_10k \
2 --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:
1uv run scripts/serve_policy.py policy:checkpoint \
2 --policy.config=pi05_franka_5task_cup_full \
3 --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
