Co-Founder & CTO @ Dirac Robotics | Brain-inspired AI Research (Compositional SSL) @ Lossfunk
Bangalore
Joined July 2012
- Tweets277
- Following502
- Followers510
- Likes2.4K
Teleoperating an SO-101 arm is such a great hand/wrist workout. My fingers are gonna get ripped by the end of the year
6
312
Physics-accurate video generation is a crucial problem to solve if we want to arrive at world models that are actually useful. The key issue is that typical diffusion-based video models don't replicate real-world physics accurately because they simply don't have the right inductive biases to "fit" the exact underlying dynamics of the scene.
We at @DiracRobotics took this challenge head on, and here's our v0 of WorldMatter, a physics-accurate gaussian simulator!
Given a few training samples of a particular type of physical interaction (eg: bowling ball hitting some pins), it's able to learn, and generalize the dynamics to give us accurate rollouts of the scene.
Check out our results in this video. There are definitely still some imperfections, but our approach is getting iteratively better with each experiment we execute.
Stay tuned for our upcoming releases!
The first step towards building a world where robots can learn.
Meet WorldMatter v0 from Dirac.
Our 3D Gaussian world model rolls out world interactions after seeing a few initial frames.
The same rollout can be observed from novel viewpoints, letting us see the interaction from different angles while preserving 3D geometry and plausible physics.
It can generalise across different motions of the world, including complex multi-collision environments.
This is just the beginning.
We've massively upgraded our Real2Sim pipeline!
Check out a detailed comparison between our pipeline, Astra and Argon. You can see that the frontier models always tend to miss out on crucial geometric and textural details that we capture accurately.
What does it actually take to turn a real-world environment into a simulation that robots can use?
At @DiracRobotics we’ve been reconstructing a Chipotle restaurant, from the overall layout and furniture down to individual objects and lighting.
Here’s a closer look. 1/5
22
566
Harsha Bommana retweeted
Fifth partnership for MOSS the open source litter-picking robot: Dirac Robotics.
Dirac Robotics is a simulation company backed by EF (Entrepreneurs First). They build a physics-accurate simulation around your robot and the place where it works, adapted to your teleop data, so you can train, replay what failed in the field and catch regressions before the next real run.
For MOSS, they are building a precise, calibrated simulation of the rover: a proper model of the body, the tracks and the arm, made from the CAD and the measured masses. Today MOSS lives in a basic MuJoCo model. With Dirac, picking up a can can be tested in simulation before it is tested on the floor, and the model stays open for anyone who builds one.
Thanks to the Dirac team for bringing serious simulation to a small litter-picking robot.
Harsha Bommana retweeted
6 months ago, we had an idea.
Now, @DiracRobotics has raised its pre-seed from @joinef 🚀
We started with one question:
If foundational models are going to transform robotics, how do we actually get those robots into the real world?
We’re building the simulation infrastructure to make that happen.
Our simulations are already helping some of the top robotics companies deploy faster and safer.
And we’ve assembled an absurdly talented team from CMU, IIT, Lossfunk, FAANG with some of the most notable robotics researchers as advisors.
Robotics is having its renaissance.
We’re just getting started. ⚡️
@harshablast, @akashizzle, @MokshitNaidu, @Alicebentinck, @rahulsamat
Harsha Bommana retweeted
A quant at an HFT firm and a retail trader can have the same view on short-term price movement of bitcoin but the retail trader can't monetize it.
On average the absolute price change in BTC ~$40 in 60 seconds.
The retail trader's round-trip fee on a leading exchange like Binance or Hyperliquid is ~$70-$80 for 1 BTC perp notional (given BTC price today).
The quant firm's VIP fee tier means their round-trip fee is between $0 (maker) and ~$30 (taker).
So for the SAME prediction, the quant firm prints money while the retail trader LOSES money after fees.
An exchange for short-term binaries that pays out on direction and charges the same fees to everyone levels the playing field.
+ You get paid for predicting the direction right, not being fastest.
+ You trade against other traders on the exchange, not the house/B-book.
+ You pay the same fee as everyone else; NO EXCEPTIONS.
Welcome to @omnibook_xyz
Absolute legends! Super bullish on @archgen_ 🔥🚀
@HariAyapps and @naveen_venk gave one of the most technically dense talks at @lossfunk
Super excited to see what's next from them!!
There are two primary ways to train your robot; I) Imitation Learning, II) Reinforcement Learning.
The former needs a lot of data, either teleop or human (with a rig), which is expensive and unscalable. Plus IL tends to be a brittle training approach.
The latter requires something harder - physics-accurate simulation environments.
We don't really have any way to create an accurate sim environment automatically, we usually have to go through the manual 3D/Sim modelling route, or give up and use assets from omniverse or lightwheel.
Today, that changes.
We can help you get YOUR assets, YOUR scene, YOUR physics, into a sim in the fastest and easiest way possible. Now you can start training your robot in your specific environments as soon as possible.
This is what we're building at Dirac Robotics. Robotics has never seen simplicity like this before!
Let us know if there are any assets or scenes you need made! We'll get it done.
Building this with @divyans1461 , @joinef
#robotics #physicalai
Le(s)go
Working on physics simulations has permanently damaged my brain.
I don’t see building blocks, I see contact friction.
#robotics #simulations #real2sim
I've been building robots since the 4th standard.
It's hard. Things break at every level. There's so much uncertainty. It takes months to get basic functionality in place. There are countless sleepless nights where we wonder if the deployment will work safely.
But, what if we could make it easier? That's what we're aiming for at Dirac Robotics. We're building crucial infrastructural pieces to actually enable robotics deployments.
So happy to be building this alongside @divyans1461. Robotics community, this one's for you.
Left Microsoft. Shut down my last company. Turned down CMU.
All of it for one problem in robotics.
Four months later, @harshablast and I solved it. Coolest thing I've ever built.
This one's for the robotics community. Launching today.
#robotics #sim2real
Glad you enjoyed the session!
Incredible session on Cognitive Science x AI at @lossfunk
No surprises about the insane knowledge room. Hoping to attend more of these.
1
15
2,076
Super excited for these folks! LFGG 🔥🚀
Introducing Agnost AI (YC S26): Product Analytics for AI Agents.
Analytics was never built for chat & voice. And that is the new interface. Your users' hidden feature requests, bugs & churn hide in conversations with your agent.
Analyzing 1M+ daily. Yours are still unread.
2
6
1,229
Education is something that is personally very close to me. What's going on right now is horrible and pathetic
There's absolutely nothing wrong with fighting for our basic rights. Even after all this, if we can't even establish basic accountability, then we're in a really dangerous situation as a country
So proud of our youth! They really showed tf up 🔥💪
5
793
The pipeline outputs more realistic scenes now! Still have a lot of work to do, will render some cool robot demos to show the physics-realism
An LLM has “attention.” It can “remember.” It appears to learn “concepts” and “generalise” from experience.
But when cognitive science and AI use the same words, are they describing similar processes—or are we hiding fundamentally different mechanisms behind shared vocabulary?
Contemporary AI is often described using precisely the same language. Yet the similarities can be illuminating in some cases and deeply misleading in others.
On July 25, I’m leading an AI × Cognitive Science research mixer at @lossfunk in Bengaluru.
We’ll begin with an accessible introduction to foundational ideas from cognitive psychology, directly compare human cognition with contemporary AI, and then open the room for a rigorous discussion:
- Where does the analogy hold?
- Where does it break?
- And what might cognitive science teach us about building and understanding more capable AI systems?
This is an intimate, topic-first roundtable—not a conventional networking event. We’re looking for researchers, engineers, writers, philosophers, artists, and independent tinkerers with strong research instincts.
Formal credentials aren’t the filter. Depth of curiosity is.
If this sounds like your kind of room, apply here: luma.com/v5p3cx1h
Thanks so much @sukanthoriginal and @scared_ape for making this happen!
Harsha Bommana retweeted
1/ A protein is a nanomachine - moving parts, does real work, changes shape to do its job. A gene editor is one of them: a molecular machine that hunts down a stretch of DNA and cuts it. We still borrow these machines from nature and tune them, because we don't understand them well enough to build our own. We started Mandrake to change that.
Harsha Bommana retweeted
beginning of the coolest research mixer in town.
drop by if world models tickle your brain.
luma.com/91t6b79g
in company of the greats @sukanthoriginal @whyloop_ @NarendraAironi @dhruvtrehan9 @vitransformer @adithya_s_k @navbenny @MayankGoel28
