Nadeesha Amarasinghe

278 posts

Nadeesha Amarasinghe

Nadeesha Amarasinghe

@nadeesha99

Machine Learning @sundayrobotics Prev: Machine Learning + ML Systems @Tesla_AI, @Apple, @Nvidia. Learned stuff at @UofT.

San Francisco, CA Katılım Aralık 2013
384 Takip Edilen4.6K Takipçiler
Sabitlenmiş Tweet
Nadeesha Amarasinghe
Nadeesha Amarasinghe@nadeesha99·
Excited to share I’ve joined @sundayrobotics The last 7 years at @Tesla_AI have been the most exciting and impactful journey of my career. I’ve had the opportunity to build and lead the ML systems that now power FSD and Optimus. From large scale training of video -> action models, research tooling, data infra, quantization for efficient inference on HW3/4/5 and evaluation. In this time we transitioned from a mostly explicit stack (v9) to one that is e2e (v14) this took lots of iteration and optimization over the entire stack by an incredible team! Today FSD is singular as an AI system that acts in the real world, at massive scale (millions of cars and billions of miles) and generalizes across multiple environments. For anyone that’s driven V14 it’s clear we are on the cusp of a major change and I’m grateful to have played a part in this. Why @sundayrobotics ? - Progress in robot learning is accelerating and @tonyzzhao @chichengcc have genuinely pushed the field forward (Aloha, Diffusion Policy, UMI, Action Chunking...) - Data (high quality, diversity and quantity) is the foundation of any great ML system and we've built a tremendous data engine - HW/SW/ML co-design -> our robot, data collection and software infrastructure are all co-optimized for scaling robot learning - Most importantly a focused, strong and full-stack team with some of my favorite @Tesla_AI colleagues @nishantsdesai @perryzjia We are at the beginning of an exciting exponential in physical AI and this team is uniquely positioned to lead it. If this excites you consider joining sunday.ai/careers
Tony Zhao@tonyzzhao

Today, we present a step-change in robotic AI @sundayrobotics. Introducing ACT-1: A frontier robot foundation model trained on zero robot data. - Ultra long-horizon tasks - Zero-shot generalization - Advanced dexterity 🧵->

English
33
23
594
178.6K
Nadeesha Amarasinghe
Nadeesha Amarasinghe@nadeesha99·
What it actually takes to build and ship robots to the wild -> full-stack systems thinking and lots "bits" and "atoms" engineering. sunday.ai/careers
Oier Mees@oier_mees

If you missed @chichengcc's guest lecture on "Robotics: Beyond Algorithms" from my @ETH robot learning course, check it out on YouTube! He shares insights that are rarely taught & hard to learn in academia. 📽️ YouTube: youtu.be/tvFvIEOBKfM 📚 Course: cvg.ethz.ch/lectures/Robot…

English
1
3
37
3K
Daniil Medvedev
Daniil Medvedev@DaniilMedwed·
Hi @united…need a little help. Flew from PSP to Florida yesterday and none of my bags arrived. Kind of need them to play in the @MiamiOpen 😉….can you help?
English
370
966
20.7K
2.1M
Nadeesha Amarasinghe retweetledi
TBPN
TBPN@tbpn·
Sunday Robotics' @tonyzzhao says they're using gloves worn by people doing household chores to create the dataset they'll use to train their foundation model: “This gives us really high-quality data, but also a really high diversity and quantity of data.”
Tony Zhao@tonyzzhao

Skill Capture Glove aligns the hands, but what about the rest of the body? Human collectors vary in height and arm length, and are also visually different. We developed Skill Transform, a method that converts glove data into equivalent robot data with a 90%+ success rate.

English
6
37
230
43.8K
Nadeesha Amarasinghe retweetledi
Alper Canberk
Alper Canberk@alpercanbe·
not long ago, i had to write code by hand and push robots around for evals. nowadays, coding agents schedule my experiments while i observe massively parallel real-world evals from my desk. it's a lot of fun. unlocked by the recent funding + the team's hard work, we have accelerated the research loops further, and are now rapidly discovering the truths of embodied intelligence with an order of magnitude more data, compute, and evals. there are so many exciting results, and we can’t wait to share more soon.
Tony Zhao@tonyzzhao

We raised $165M at a $1.15B valuation to stop doing demos. 2026 is about 1) deployment and 2) research. We will start shipping Memo with our new frontier models in a few months. Our series-B is led by Coatue, with Thomas Laffont joining the board. ->🧵

English
6
17
148
17.7K
Nadeesha Amarasinghe
Nadeesha Amarasinghe@nadeesha99·
We’re going full speed on real-world deployment! Our ML iteration loop is the most innovative and impressive I’ve seen in robot learning. Expect a lot more exciting research progress this year. Join us sunday.ai/careers
Tony Zhao@tonyzzhao

We raised $165M at a $1.15B valuation to stop doing demos. 2026 is about 1) deployment and 2) research. We will start shipping Memo with our new frontier models in a few months. Our series-B is led by Coatue, with Thomas Laffont joining the board. ->🧵

English
2
3
21
2.6K
Nadeesha Amarasinghe
Nadeesha Amarasinghe@nadeesha99·
@arnie_hacker Outside of storage you must consider the additional network I/O of reading uncompressed frames, increase in host memory traffic moving uncompressed frames from dataloader workers to main process and increase in host to gpu transfer
English
0
0
2
164
Arnie Ramesh
Arnie Ramesh@arnie_hacker·
Anyone experienced with training video diffusion models? Noob question: Do you pre-process mp4 into individual frames and store before training? Doesn't this blow-up storage requirements? Or do you dynamically convert mp4 into frames during training (how is this parallelized?)
English
16
2
73
13.1K