about datesets

[lunalulu]

thanks for your great work！

The paper states that 7,500 trajectories were used from Teleoperation， but upon downloading the dataset, it was discovered that there are actually 9,500 trajectories from Teleoperation. So, should I use these 9,500 data points to train my model?

thanks for your reply~

[Jdvakil]

Hey, thank you so much for reaching out. We only used a subset (7,500 trajectories) of the 9,500 trajectories. Feel free to use all of the 9,500 data points for your model training.

[lunalulu]

THANKS~ @Jdvakil

[lunalulu]

@Jdvakil
Another question, why is the following restriction added to the data, which is equivalent to only 40% of the total data being used for training? thanks for your reply~

[Jdvakil]

This variable can range up to a maximum of 250 (total trajectories per task). We ran a series of ablation studies aimed at understanding how varying data volumes impact the agent's performance and its ability to generalize.

[lunalulu]

@Jdvakil thanks, i get it~

[Jdvakil]

@lunalulu btw, we ran the studies for 50, 100, 150, 200, 250 demonstrations

[lunalulu]

@Jdvakil Aren’t the more data the better?

---

Another question, I didn’t find the semantic Augmentations part and parameter settings, can you tell me?

Or is there no semantic Augmentations part in the project?

thanks for You~

[Jdvakil]

@Jdvakil Aren’t the more data the better?

We wanted to demonstrate that our agent can perform and generalize over multiple scenes and configurations without depending heavily on a plethora of data.

[homangab]

Thanks for the question - the repo currently doesn't have the semantic augmentation scripts. We will be adding them soon (in about 3-4 weeks!)

[HM102]

@homangab any plans to still add it?

[leslieburke]

@homangab any plans to update?

[homangab]

Apologies for not posting the update here - we were not able to obtain the required approvals from Meta for releasing the generative model based augmentation code due to some compliance rules. I'll close this issue now, feel free to re-open if there is anything else. Thanks!