openpi

openpi copied to clipboard

[Feature request] Add or document “subtask prediction” stage (π0.5 two-stage inference)

6

- **Summary** The π0.5 paper describes a two-stage inference procedure: first generate a high-level subtask text (e.g., “pick up the pillow”), then predict low-level actions conditioned on that subtask. In...

ajs7270

Finetuning on 7-DoF arm. which dataset format should I use?

1

I need to finetune pi0 on my own 7-DOF + gripper dual arms. I notice that aloha dataset is based on 6-DoF arms while Droid dataset is based on 7-DoF...

lin-whale

CPU Memory Leak When Saving Checkpoints During Fine-tuning

3

**Description** I've observed a progressive increase in CPU memory usage during the fine-tuning process. Specifically, after each model checkpoint is saved, a portion of the CPU RAM is not released....

tingjun-cs

Gripper close too early

6

Hi Author! Thank you for taking the time to read this issue. I'm currently doing simulation using pybullet. My task it to pick up a white cube and put it...

reinaldoyang

Hello, while studying the Pi0.5 model, I have two questions regarding the model implementation that I would like to ask you: 1、The paper mentions that the model adopts two-stage pre-training...

Ginldaj

Asking very nicely

sean1295

https://github.com/Physical-Intelligence/openpi/blob/175f89c31d1b2631a8ff3b678768f17489c5ead4/src/openpi/models/model.py#L243 ```py def load_pytorch(self, train_config, weight_path: str): logger.info(f"train_config: {train_config}") model = pi0_pytorch.PI0Pytorch(config=train_config.model) safetensors.torch.load_model(model, weight_path) return model ``` In this code, there is no way to set [dtype](https://github.com/Physical-Intelligence/openpi/blob/175f89c31d1b2631a8ff3b678768f17489c5ead4/src/openpi/models/pi0_config.py#L20) like trainig script...

icsl-Jeon

Finetuning on 7-DoF arm. which dataset format should I use?

1

I need to finetune pi0 on my own 7-DOF + gripper dual arms. I notice that aloha dataset is based on 6-DoF arms while Droid dataset is based on 7-DoF...

lin-whale

Hello, Thanks for the excellent work! @kpertsch: I am curious about the processing of state in FAST related experiments: prompt is concatenated with discretized state tokens to form a new...

xianjiedai

Is PI0.5 finetuned by all LIBERO datasets? Success rate of LIBERO-90 is only 18%?

4

libero_spatial | 97.4% libero_object | 98.4% libero_goal | 97.6% libero_10 | 93.0%

Kimho666