Isaac Tu

Founding Research Engineer @ Human Archive; NYU GRAIL Alumni

[Isaac Tu @Isaac1385]

@pushpak_agr @babugi28 @microagi @MeckaAI yeah we would need thousands of trials to get something statistically significant. We never claimed statistical significance. We are simply showing our results from our experiments and showing the competitiveness of Human Archive data.

[Isaac Tu @Isaac1385]

@pushpak_agr @babugi28 @microagi @MeckaAI Sampling was random episodes from egoverse. 100 hours is the same order of size as VITRA's own pretraining dataset. training was identical for each arm, 20k steps at batch 512, around 1 pass over each 100h.

[Isaac Tu @Isaac1385]

@bercankilic @babugi28 @microagi @MeckaAI gotcha. we just worked with what was publicly available on egoverse and took an honest sample from that.

[Isaac Tu @Isaac1385]

@bercankilic @babugi28 @microagi @MeckaAI we took a random sample of episodes, sampled from all 2000 hrs.

[Isaac Tu @Isaac1385]

@Deepans36819800 @babugi28 @microagi @MeckaAI why do you say that?

[Isaac Tu @Isaac1385]

Big things coming soon! We are getting better every release!

[Raj Patel @babugi28]

We pre-trained the same VLA on 100 hours of Human Archive’s OTS dataset and 100 hours of randomly sampled data from each of @microagi’s and @MeckaAI’s datasets on EgoVerse.

Overall, Human Archive data achieved the strongest results, with the highest overall task success rate. We observed particularly strong generalization to unseen objects.

Thread below 🧵👇

[Isaac Tu @Isaac1385]

@babugi28 @microagi @MeckaAI Big things coming!

[Isaac Tu @Isaac1385]

@MilesDigitek @babugi28 wasn't a calibration mismatch. regarding viewpoint, are you saying the tasks are ood of the pretraining data viewpoint?

[Isaac Tu @Isaac1385]

An early look at what our data does for downstream robot manipulation. More coming soon

[Raj Patel @babugi28]

We benchmarked only 7 hours of Human Archive data against 4 widely used public datasets totaling 17 hours: HOT3D, H2O, HOI4D, and TACO.

The results:

Human Archive only: 85% task success | 65.6% task progress
4 public datasets: 57.5% task success | 52.5% task progress
Mixture of all 5: 60% task success | 56.3% task progress

[Isaac Tu @Isaac1385]

@babugi28 Enjoyed helping out evaluating our data on a robot for the first time! Much more coming soon!

[Raj Patel @babugi28]

We benchmarked only 7 hours of Human Archive data against 4 widely used public datasets totaling 17 hours: HOT3D, H2O, HOI4D, and TACO.

The results:

Human Archive only: 85% task success | 65.6% task progress
4 public datasets: 57.5% task success | 52.5% task progress
Mixture of all 5: 60% task success | 56.3% task progress

[Isaac Tu @Isaac1385]

@babugi28 bro why did you post this lmao

[Isaac Tu @Isaac1385]

@babugi28 let's go!!