ahad

966 posts

ahad banner
ahad

ahad

@ahadj0

building autonomous robots | prev @aws @amazon

Katılım Ekim 2022
172 Takip Edilen846 Takipçiler
Vector Wang
Vector Wang@VectorWang2·
Why can't “world models” just be sim? A real-time real2sim demo in the upcoming “ManiDreams” Foundation Stereo + SAM2 + Newton = Generalizable 3D physics prediction in a fast (15FPS) and lightweight (3070) way. No data or object mesh needed rice-robotpi-lab.github.io/ManiDreams/
English
3
2
13
378
ahad
ahad@ahadj0·
@chris_j_paxton it interesting to see the difference of public funding and private funding for ai research
English
0
0
1
121
Chris Paxton
Chris Paxton@chris_j_paxton·
I do increasingly think that open-source ai is of fundamental importance to human liberty, and it's worrying how under-emphasized it is outside of, like, literally just AI2
English
10
9
86
7.2K
ahad
ahad@ahadj0·
@PrathamJainAI so, when I trained the reward model on the wrist view it actually was having a pretty hard time fitting the data
English
0
0
0
20
Pratham Jain
Pratham Jain@PrathamJainAI·
@ahadj0 For my setup I use 3 cams one at top one at the wrist of the robot and third front view
English
1
0
1
21
ahad
ahad@ahadj0·
@PrathamJainAI is there a particular angle you think could help?
English
1
0
1
31
Pratham Jain
Pratham Jain@PrathamJainAI·
@ahadj0 Try another camera and add it in front view and let magic happen
English
1
0
1
28
ahad
ahad@ahadj0·
@PrathamJainAI interesting observation maybe including more lighting could make it robust to this. did you see some correlation between shadows and predicted value?
English
3
0
1
36
Pratham Jain
Pratham Jain@PrathamJainAI·
@ahadj0 You also have bad lighting avoid shadows formation this creates confusion in the model
English
1
0
1
31
Karan Jagtiani
Karan Jagtiani@karanjagtiani04·
@ahadj0 Curious to see the impact of removing the wrist view. Any initial results to share?
English
1
0
1
53
ahad
ahad@ahadj0·
interesting to see which camera view fits the best for a reward model trained on 26 mins of data this is inference on a demo in the training set
English
3
5
56
5.3K
ahad
ahad@ahadj0·
@dappsidesanchez seems like the wrist views were actually getting in the way so trying to train the reward model without the wrist view and going to try training a policy with it
English
0
0
0
56
Sanchez | dappside.io
Sanchez | dappside.io@dappsidesanchez·
@ahadj0 Can’t wait to see how the view affects long term training. keep em comin!
English
1
0
1
70
ahad
ahad@ahadj0·
@iamRezaSayar not atm maybe after the research is complete
English
0
0
1
141
Reza Sayar
Reza Sayar@iamRezaSayar·
@ahadj0 nice! was this your idea? planning on open sourcing it?
English
1
0
1
181
ahad
ahad@ahadj0·
@RealJoshuaYang @PrathamJainAI i'm not the author of that tweet so i can't really answer for them but in my experience using so 101 arms you would need a third person camera + wrist camera
English
1
0
1
29
Joshua Yang
Joshua Yang@RealJoshuaYang·
@ahadj0 @PrathamJainAI Hmm ok I have a 3090 locally, what hardware are using for inference in that video? Also is your setup wrist camera + birds eye camera?
English
2
0
1
26
ahad
ahad@ahadj0·
@chris_j_paxton also something to note when ever your record anything with lerobot (huggingface's robot) library they by default upload it
English
0
0
0
87
Chris Paxton
Chris Paxton@chris_j_paxton·
The robotics data wall is falling
Pierre-Alexandre Balland@pa_balland

Something big is happening in robotics - and it’s hiding in plain sight. This post is not about dancing robots but in the data that powers them. Open robotics datasets have exploded this year, turning the field into a more scalable and collaborative ecosystem. In just two years, @huggingface datasets grew from 11k to over 600k - and robotics is by far the fastest-growing segment. We went from 1k robotics datasets in 2024 to 27k in 2025! For comparison, text generation, the second-largest category, has only around 5k datasets in 2025. That gap is massive. Open datasets are important because robotics lives and dies by real-world robot data - video, actions, sensors, failures. By making this data easy to upload, reuse, and benchmark, researchers, startups, and large players are now releasing real-robot datasets that would have stayed locked inside labs just a few years ago. Major contributors include @nvidia, LeRobot initiative, and a rapidly growing maker community. This surge is also enabled by cheaper video storage, better tooling, and an open-source AI culture now spilling into the physical world. And it really matters: open robotics data dramatically lowers entry barriers, accelerates learning-by-doing, and speeds up progress toward generalist and humanoid robots. Robotics won’t scale through hardware alone - but to a large extent through shared data. Viz below from @aiworld_eu - link to the story and more viz/filters in comment.

English
6
31
319
37K
Joshua Yang
Joshua Yang@RealJoshuaYang·
@ahadj0 @PrathamJainAI By the way, do you think diffusion policy would be feasible on the SO Arm 101 arms? What arms are you using in the video?
English
1
0
1
30
ahad
ahad@ahadj0·
@kamalgupta09 The way the PI 0.5 paper described it, they trained their VLM backbone on a mixture of discrete action tokens and multimodal web data.
English
0
0
0
138
Kamal Gupta
Kamal Gupta@kamalgupta09·
Had a few of conversations during NeurIPS where the researcher think of VLAs as VLMs whose autoregressive head is fine-tuned to output actions tokens. Is that the generally accepted definition? I always thought of them as any model trained from mixture of data that has understanding of vision, language, and actions (irrespective of whether or not if it’s autoregressive, or if it has internet pretraining)
English
8
0
45
5.5K
Andreas Köpf
Andreas Köpf@neurosp1ke·
"changes everything" ablated by PI ;-) - seems yes, wrist cams are overall nice to have but not all tasks benefit...
Andreas Köpf tweet media
English
2
2
26
4.1K
ahad
ahad@ahadj0·
@d_ol_an @LeRobotHF You don't need 2 so 100 robots you can get away with 1 if you use some other way of teleoping but it is similar if you use 1 so100 follower and 1 leader
English
1
0
0
23
ahad
ahad@ahadj0·
here’s a demo of removing and attaching the gripper by itself robots that can use any tools!
English
24
35
462
27.2K