How much data is needed for model fine-tuning when testing with the G1 robot in simple scenarios?

[nooetic]

For example, how much data is needed for scenarios like picking and placing objects or pouring water.
The pouring water scenario used 300 data samples for training, but the real robot performed poorly in closed-loop testing, while it worked well in open-loop testing.

[ModiShi]

100 episodes should be enough for pick & place tasks, while more diffcult tasks like poouring water may need more. However, the performance of the model is also related to the quality of your collected data.

[nooetic]

the performance of the model is also related to the quality of your collected data

How to evaluate the quality of a dataset

[ModiShi]

It's difficult to quatify the data quality, but our experience suggests prioritizing motion pattern consistency and minimizing redundant or non-useful actions would be beneficial.

[nooetic]

It's difficult to quatify the data quality, but our experience suggests prioritizing motion pattern consistency and minimizing redundant or non-useful actions would be beneficial.

Thanks for taking the time to explain that.