Shreyas Gite (@shreyasgite) / X

Shreyas Gite

1,441 posts

Shreyas Gite

@shreyasgite

Mainly dad, building at tuul.ai | Prev: self-driving founding-eng @Ford acq., jet-engines @RollsRoyce, oil-rigs Transocean

Berlin, Germany

tuul.ai

Joined September 2009

Pinned
Shreyas Gite
@shreyasgite
Mar 21, 2025
Gemini-powered robot can now effectively debug itself! I've been obsessed with two main questions in robotics: can robots learn from their own mistakes without humans in the loop, and how much can we leverage synthetic data? Spoiler: yes, and it's surprisingly elegant once you
00:00
47K
Shreyas Gite
@shreyasgite
Oct 20, 2025
Karpathy manually went through every training image on ImageNet. Then the entire test set. Just to set a human benchmark. He recently implemented a 1989 LeCun paper to measure how much efficiency you gain by simply knowing the deep learning future. And these are just two examples
545K
Shreyas Gite
@shreyasgite
May 15, 2025
Gemini + π0 = actually useful robots! (Similar to what @physical_int did with "Hi Robot") I can now verbally tell the robot that I'm building a red Lego wall or wooden tower, and it will infer the next steps by itself and pass me the necessary pieces, tools, or materials, ha!
00:00
25K
Shreyas Gite
@shreyasgite
Apr 9, 2025
Not many know that Gemini 2.5 or Flash can do zero-shot embodied reasoning (ER) for tasks like planning trajectories, grasping points, 3D boxes etc. Which is good enough for most research & experimentation. So if you in the waitlist and still haven't had access to official
01:42
Norman Di Palo
@normandipalo
Apr 7, 2025
Last month we announced Gemini Robotics (GR) and Gemini Robotics-ER (GR-ER). GR-ER is a powerful VLM specialised for spatial understanding, including detecting object poses in 2D/3D, pointing, and even *predicting grasp poses*. Take a look at this demo. Details below. 🧵
29K
Shreyas Gite
@shreyasgite
Apr 8, 2025
🤖 Diving into Robot Foundation Models! 🧠 Starting a small series of Demystifying Robotic Foundational models starting with πο (Pi0) from @physical_int. Understanding Pi0 gives great context for later models like Gemini Robotics & GR00T N1. The core idea behind many robot
00:00
Karol Hausman
@hausman_k
Feb 4, 2025
We're open-sourcing π₀! Making physical intelligence a little more real and accessible to all 🦾🤖 blog: pi.website/blog/openpi repo: github.com/Physical-Intel…
24K
Shreyas Gite
@shreyasgite
Aug 22, 2025
Some interesting bits from the Boston Dynamics demo and post: - Humanoids can, in principle, tackle a tremendous breadth of manipulation tasks, so actually quite a good form factor for general robots. - Collecting data of robots recovering from disturbances is quite important to
00:00
47K
Shreyas Gite
@shreyasgite
Sep 8, 2025
Having seen so many laundry folding robot demos, I started training with my bimanual setup, and gotta say I have new respect for those companies and individuals. Even collecting 200 episodes is tricky, and I cannot easily (yet!) bootstrap learning with affordances like I was able
00:00
36K
Shreyas Gite
@shreyasgite
Oct 15, 2025
Data collection is a high-value task. Even the Chancellor of Germany has to do his part. Friedrich Merz with @LeRobotHF SO100.
65K
Shreyas Gite
@shreyasgite
May 13, 2025
Pi0 vs. ACT with BBox conditioning 🟦 Not many know you can push ACT to *almost-Pi0* generalisation by conditioning on bounding boxes (BBoxes). How is the training data collected? • Generate BBoxes for all pick-and-place objects in the scene.(I used Gemini) • Pick-and-place
00:00
00:00
35K
Shreyas Gite
@shreyasgite
Jun 25, 2025
Working on a fork of Aloha Mujoco Sim from awesome Gemini-Robotics team for @LeRobotHF so100/so101. Hopefully soon 🦾
00:00
43K
Shreyas Gite
@shreyasgite
Feb 8, 2025
🚀 Google Gemini on $120 robot arm to tidy up my kid’s room! Sorting Legos is painful, so I teamed up with my son to train a robot arm. @LeRobotHF from @huggingface works like charm to train a policy. However I wanted to try LLM end-to-end control—image in, motor values out.
00:00
42K
Shreyas Gite
@shreyasgite
Jun 17, 2025
If you are thinking about world models or neural sims and don't know where to start, check out the OG paper on world models from @hardmaru and @SchmidhuberAI - It is super clear to read and get the basics - you can reproduce it on your Mac or any local machine - you can steadily
00:00
14K
Shreyas Gite
@shreyasgite
Nov 13, 2025
Most robotics startups don’t even know what they’re missing. Tesla is absurdly far ahead in robotics: • A quasi VLA that works more like an omni-model, plus a reasoning-driven that can navigate genuinely tricky situations • Generating tons of data isn't the solution; finding
Ashok Elluswamy
@aelluswamy
Oct 24, 2025
Replying to @aelluswamy
Full video of the ICCV '25 presentation
00:00
23K
Shreyas Gite
@shreyasgite
Jun 11, 2025
Learning from Human Demonstrations: Show the Robot How to Act! The pipeline is very similar to older experiments using Gemini & pi0 with @LeRobotHF. Pi-zero runs locally, while Gemini Flash generates the affordances and the high-level task. (More details are in the thread.) The
00:00
23K