π0.5 Dual-Franka Five-Task Full SFT — 20k
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
| Task ID | Training prompt |
|---|---|
place_bowl_on_cabinet |
Place the green bowl on top of the cabinet with the left arm. |
place_bowl_on_plate_pt2 |
Place the green bowl on the blue plate with the left arm. |
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
- Raw episodes: 250 (50 per task)
- Training split: 225 episodes (45 per task)
- Held-out split: 25 episodes (5 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.
Data-quality note
For place_bowl_on_plate_pt2, 32 trajectories had chest-camera timestamp offsets above the strict default conversion threshold. They were retained with a 3000 ms camera-alignment tolerance to satisfy the experiment requirement of using all 50 trajectories. This can introduce observation/action temporal mismatch and should be considered when interpreting real-robot behavior.
Training configuration
| Setting | Value |
|---|---|
| Base checkpoint | Official OpenPI pi05_base |
| Method | Full-parameter SFT |
| Training steps | 20,000 |
| Hardware | 1 × NVIDIA A100 80GB |
| Per-device / global batch | 2 / 2 |
| 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 | 1,000 steps |
| Loss logging | Every 100 steps |
| Final checkpoint index | 19999 |
The last logged training loss at step 19,900 was 0.0096. Training loss is not a real-robot success metric.
Download and loading
hf download THU98/pi05_franka_5task_full_20k \
--local-dir ./pi05_franka_5task_full_20k
The repository is an OpenPI JAX/Orbax checkpoint. Use it with a matching local OpenPI configuration named pi05_franka_5task_pt2_full:
uv run scripts/serve_policy.py policy:checkpoint \
--policy.config=pi05_franka_5task_pt2_full \
--policy.dir=./pi05_franka_5task_full_20k
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 25 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, camera calibration, viewpoints, objects, or language paraphrases.
- The five tasks are narrow and mostly left-arm behaviors; this is not a general-purpose robot policy.
- 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. The model must not directly bypass robot safety systems.
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。训练集为 225 条、留出集为 25 条，总训练 20,000 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。
place_bowl_on_plate_pt2 中部分轨迹存在较大的胸部相机时间戳偏差。本实验为保留全部 50 条数据放宽了对齐阈值，因此部署时需要重点检查视觉与动作的时间同步。当前没有发布正式真机成功率，训练 loss 不能替代真机评测。