Kirill Danilyuk

Head of CV @AvrideAI. Founded and scaled the computer vision group @Yandex SDC. Building perception robots can actually trust. Private pilot.

[Kirill Danilyuk @kd3d_ai]

@Scobleizer Two modalities are great if you can handle the technical challenges they introduce. Quick reaction is not a pure perception problem, it is also a sensor-to-control latency problem and optimizing this is frankly a big deal. We need to talk more about latency in AVs.

[Kirill Danilyuk @kd3d_ai]

It's funny that @karpathy isn't updating https://t.co/ZdwbTLdyL4 anymore. Last featured talk is from 2025 and the bio stops at 2024. Too busy shipping :)

[Kirill Danilyuk @kd3d_ai]

@elonmusk Cybercab with no mirrors, either, which might be even cooler.

[Kirill Danilyuk @kd3d_ai]

This is what the future should be like - thanks to @Starlink. It literally improves my day for free while being completely commercial at the same time.

A communist mind cannot comprehend this.

[Kirill Danilyuk @kd3d_ai]

Very cool indeed, but let's not forget the elephant in room is manipulation and spatial understanding, not the looks.

[Genesis AI @gs_ai_]

Humanoid robots don't need to look human.

Meet Eno, our first general-purpose robot.

Not a machine pretending to be human, but intelligence given a body.

At Genesis, we’re building a future where robots don’t feel cold or distant, but capable, calm, and ready to help.

Available Q4 this year.

[Kirill Danilyuk @kd3d_ai]

Cannot agree with @Nick_Davidov more. America is an anomaly: you can come here, do your best work, make a lot of money, and KEEP it. You guys who were born here just take it for granted, but that's not how the world works in general.

[Nick Davidov @Nick_Davidov]

Idc about all morons saying that society shouldn’t have trillionaires. I was born in USSR and grew up in Russia in the 90s. For me it’s a true marvel that:
- an engineer can get rich by being a good engineer
- an entrepreneur can get rich by building something people want and not grifting on the government or natural resources
- an entrepreneur engineer can become a trillionaire and not be in prison

[Kirill Danilyuk @kd3d_ai]

@Nick_Davidov Moreover, if someone became a billionaire, the pie just got bigger and there's a chance they start investing into startups I can build!

[Kirill Danilyuk @kd3d_ai]

Posts like this is absolute gold and the reason I love this platform. Must read for any hardware founder. I'm not going to summarize, go ahead and read the whole thing.

[Rui Xu @xu545302]

https://t.co/GBUEYqSj7J

[Kirill Danilyuk @kd3d_ai]

I won't believe in self-driving cars until they ditch the mirrors.

[Elon Musk @elonmusk]

Cybercab driving itself out of the GigaTexas factory

[Kirill Danilyuk @kd3d_ai]

@typesfast talks about global trade in a way that's even more engaging than his takes on making money. Genuinely fascinating stuff. Would love to hear him cover the Hansa one day.

[Ti Morse @ti_morse]

My first interview with Ryan Petersen (@typesfast), Founder of Flexport.

0:03 Pablo Escobar was a logistics guy
2:29 The explosion in tariff fraud
8:58 The Dutch East India Company
11:20 History of global trade
14:39 1,000x spice markup
17:53 The British East India Company
24:02 How the British got 20% of China addicted to opium
27:44 The Forbes family & opium trade
30:40 Jewish trading networks
37:33 It’s illegal to criticize the King of Thailand
38:58 Strait of Hormuz
45:58 Maritime chokepoints

[Genesis AI @gs_ai_]

We are back again :) After three weeks of quiet building.

Introducing Genesis World 1.0, our latest simulation platform, the second release in our full-stack suite. Open-sourced.

Robotics is still bottlenecked by the 1× speed of the physical world. Every model, checkpoint, and data recipe eventually needs to be tested on physical hardware, slowly, expensively, and with limited coverage.

One hour in reality can become 100 days in simulation. That is how robotics model iteration moves from a wall-clock bottleneck to a compute problem.

To make this work, simulation has to be both fast and trustworthy.

Over the past year, we rebuilt the entire stack: a GPU-accelerated cross-platform compiler, penetration-free multi-physics contact solvers, unified rigid and deformable physics, and a photo-realistic renderer purpose-built for physical AI applications.

We built Nyx, a high-performance path-traced rendering engine for robotics application.

Genesis World 1.0 achieves near realtime performance with our latest development for penetration-free IPC solver, supporting various types of deformables beyond rigid bodies. It supports contact-rich, dexterous manipulation simulation across different embodiments: unitree, sharpa, wuji, genesis hand and various types of grippers.

Under the hood is Quadrants, our effort in pushing forward cross-platform GPU-accelerated computation. Quadrants started as a fork of Taichi, and we rebuilt most of the critical parts for optimizing simulation workloads, giving 10x faster launch time and up to 4.6x runtime performance compared to the initial Genesis release.

Together, they bring us to an unprecedentedly low sim-to-real gap, enabling zero-shot real-to-sim model evaluation and much faster iteration of GENE.

All available today.
Genesis World 1.0: https://t.co/aknCM3eqws
Quadrants: https://t.co/uXqPNI4cb6
Nyx: https://t.co/R8j0djqGnV

[Kirill Danilyuk @kd3d_ai]

Wow, this Wednesday keeps on giving. This would be super interesting to dig into.

[Felix Heide @_FelixHeide_]

Excited to introduce #TruckDrive 🚛 at #CVPR2026: a new long-range driving dataset built specifically for long-range truck autonomy, where safe braking and anticipatory planning demand perception hundreds of meters ahead, far beyond existing robotaxi datasets.

📦 TruckDrive includes:
🔹 475K samples, with 165K densely annotated frames
🔹 Benchmarks for end-to-end driving, tracking, planning, depth estimation, and up to 1,000m for 2D detection and 400m for 3D detection 📏🎯

🛰️ A purpose-built long-range sensor suite:
🔸 7 long-range FMCW LiDARs (range + radial velocity)
🔸 3 high-res short-range LiDARs
🔸 11× 8MP surround cameras for short and long-range📷
🔸 10× 4D FMCW radars 📡

⚠️ Key finding: current state-of-the-art models break down at long range

📉 with 31% to 99% drops on 3D perception tasks beyond 150m. TruckDrive exposes a long-range generalization gap that current architectures and training signals are not closing yet - a benchmark for the next generation of long-range highway autonomy research 🚚

🔗 Project and Data: https://t.co/fNzDCbGQRQ

Fun work together with @torc_robotics led by Filippo Ghilotti, Edoardo Palladin, Samuel Brucker, Adam Sigal, and Mario Bijelic.

[Kirill Danilyuk @kd3d_ai]

Genesis team became publicly active lately, and that's very good news for the robotics community. They introduced Genesis sim as a concept back in 2024, and then went quiet. Now they are back with https://t.co/877Zo4MQgu and Genesis World https://t.co/O7s4mpV21H

[Kirill Danilyuk @kd3d_ai]

I think people don't appreciate enough how hard this is. I've only caught a glimpse of the stream, but several important things are clearly being done right:

The robot got to the line on its own. Robots coordinate battery levels across the fleet, when one gets low, another one comes up to take over. No visible overheating, no fatal pipeline crashes from random segfaults, no memory leaks that only surface after hours of operation. Pipeline latency looks fine. Control looks solid: coordinating so many DoF is no joke. Manipulation has room for improvement, but that's understandable.

@adcock_brett , would love to see a highlights video after this ends! Things like robots walking in/out, challenging moments, out-of-distribution problems, unexpected behavior and basic stats such as accuracy. Keep it up.

[Brett Adcock @adcock_brett]

Details on what's going on:

> Our original goal was an 8-hour run - we wanted to run nonstop and fully autonomous. Since then, we made the decision to keep the party going. We’re now over 48 hours of nonstop autonomous operation without a failure to perform the use case. This is uncharted territory

> The task is small package sorting. F.03 detects the barcode, picks up the package, and reorients it barcode face-down onto the conveyor

> Humans average around 3 seconds per package. F.03 is now around human parity. The robots are reasoning directly from camera pixels in the robot head

> The robots are fully autonomous running Helix-02, our in-house neural network running entirely onboard F.03. There is no teleoperation - every action comes directly from Helix-02

> If the robot gets stuck or the AI policy goes out of distribution, Helix triggers an automatic reset. You’ll occasionally see this happen during the livestream

> YouTube commenters started naming the robots Bob, Frank, Rose, and Gary this week, so we added name tags to each robot

> If a robot has a software or hardware issue, it autonomously leaves for maintenance and another robot takes over. We run our labs at Figure this way to maximize uptime. We haven’t had a failure yet, but statistically we probably will at some point

We are now running this until a failure to perform the use case!

[Kirill Danilyuk @kd3d_ai]

Grok is actually amazing. The selling point for me is the Ara voice + argumentative personality. It pushes back hard and the voice itself is superb, it handles my Russian accent like a charm. And it feels like a real conversation in a way that's hard to describe.

[Elon Musk @elonmusk]

Grok upgrades

[Kirill Danilyuk @kd3d_ai]

There is a computer vision course I wish I had taken properly before starting out in perception: https://t.co/VPHOBkRxBO

No, it is not CS231N, the famous CNN-heavy course by @karpathy. It is CS231A, its geometry-first counterpart. You need to know the fundamentals: camera models, calibration, SfM, stereo/mono depth estimation, scene flow, etc. Of course, end-to-end transformers can learn a lot of this on their own, but at least you need to know what is going on and why.

Amazing course notes and assignments. If you complete the course and can reimplement something like LSS (Lift, Splat, Shoot), you'll be well positioned for the job regardless of current trendy architectures. Do not skip this part.

[Kirill Danilyuk @kd3d_ai]

The bottleneck for AI engineers has shifted from producing code to verifying it. And verification is getting really hard: ML bugs don't crash anything, they just silently train the wrong thing. And no, you can't catch what you don't already understand.
Years ago at Yandex I interviewed ML candidates with exactly this task: here's the training code, find the subtle bugs. Back then it was a filter for senior engineers, now it's the core of the trade. What a time to be alive.

[Kirill Danilyuk @kd3d_ai]

Progress podcast's signal to noise ratio is extremely high. And this is only episode two. Keller and Ryan are ideal role models for any serious entrepreneur: people who take real risk, build real businesses, and compound company value year after year for decades. @fuelfive, outstanding work on guest selection, please keep it up!

[D. Scott Phoenix @fuelfive]

The Silk Road made everyone rich, and then it killed half of them.

Progress ep02 is live with @typesfast of @Flexport.

We discuss why the global economy is as fragile as ever, what it takes for America to build again, and whether AI needs its own god.