Robots need tons of real examples to learn new skills. But right now, getting good training data is slow and expensive. You usually need experts and special gear. $COBALT changes that. It’s a simple app where anyone can use their regular phone to control robots from anywhere. Just move your phone and the robot follows. Helps collect lots of clean data fast and cheap. I’ve been following it for a while and it’s one of the more practical ideas I’ve seen for scaling robot learning.

🧵1/ COBALT is the future of Robotics. Built by @animesh_garg (ex-NVIDIA, Georgia Tech robotics professor), COBALT is tackling one of the biggest bottlenecks in Physical AI: Data. Not another robot company. Infrastructure for robot learning.

The real bottleneck in scaling AI robotics isn’t just better models or more powerful hardware. It’s the ability to collect high-quality, diverse robot demonstration data at scale, across many different environments and by many different operators. Right now, most robot data is still collected in controlled labs with expensive setups. But the projects that are building cloud-based systems to turn ordinary smartphones into tools for remote robot control are quietly changing that. By making teleoperation accessible and parallelizable, they’re creating the kind of distributed, real-world data infrastructure that generalist robots will eventually need to move beyond simulation and into everyday environments. This layer of infrastructure might end up being more important than most people currently realize.

One of the biggest unlocks for scaling robot learning won’t just come from bigger models or better hardware. It will come from building real infrastructure for high-quality, diverse robot demonstration data at scale, collected across many environments and by many operators. COBALT is doing exactly that: turning everyday smartphones into a global, cloud-native teleoperation system that makes it possible to gather rich robot data without expensive setups or closed labs. This kind of accessible, parallel data collection could become foundational infrastructure for the next phase of robotics. cobalt-teleop.github.io

This @YeeErc20 account is straight-up the meme overlord 😂🦖

This data tier list for humanoid & robotics foundation models is spot on. Real-world robot experience (especially high-dexterity data) sits at the top, while internet/human videos provide surprisingly strong transfer learning. Simulation and synthetic data still need more real-world proof. The missing piece for scaling high-quality robot-specific data across many environments is scalable teleoperation infrastructure. This is exactly where COBALT comes in, using everyday smartphones to collect diverse, real-world robot demonstrations at global scale. x.com/michellelsun/status/20…

Genuine question for @animesh_garg: if you had an extra $32K/month (projected on 8K first week) in unrestricted funding, what’s the first bottleneck in teleoperation research you’d throw it at? Compute? Hardware? Data labeling?

Looking at this taxonomy from the recent survey on Robot Learning from Human Videos, one thing becomes clear. Human videos are excellent for scaling observations and high-level understanding. But when it comes to grounding robot actions in the real world especially in the Action-Oriented Transfer layer (affordances & latent actions) nothing replaces high-quality, diverse robot demonstrations. The real unlock isn’t one lab producing demos in isolation. It’s scalable, cloud-native teleoperation that lets many operators across many environments generate rich robot data in parallel. This is exactly what $COBALT was designed for. Credits: arxiv.org/pdf/2604.27621

Love this framing. The line that matters most is the last one. Everything is in service of making that function good. Here's the part people underrate though. What makes it good isn't a smarter architecture. It's the demonstrations behind it. The function can only learn behaviors someone actually showed it. So the real frontier in robotics isn't the network. It's how cheaply and how widely you can collect real world examples of what you want the robot to do. Get that right and everything downstream gets easier. x.com/interlatent/status/206…

Tether wants robots that can hold a wallet and transact on their own. Fine. But a robot in the economy is only worth something if it can actually do the work. And that comes down to one thing. Training data. A humanoid with a self-custodial wallet and no skills is just an expensive paperweight with a balance. The machine economy everyone's racing to fund sits on top of capability. And capability sits on top of demonstrations. That's the layer @animesh_garg's PAIR lab is building with COBALT. Crowdsourced teleop from any phone. The cheapest way to teach robots the skills this whole economy is supposed to run on. Fund the wallets all you want. Data is the foundation. cobalt-teleop.github.io x.com/tether/status/20647290…

$1.4B into a single humanoid company. Tether Nvidia Amazon Qualcomm all in at a $7B valuation. But look closer. Neura isn't just building a robot. They're building a training data platform alongside it. That's the tell. The constraint was never the hardware. It's the data to teach it. Which is exactly what @animesh_garg's PAIR lab is attacking with COBALT. Turn any phone into a teleop rig and crowdsource demonstrations at scale. No expensive hardware. The cheapest path to the one input every humanoid actually needs. Capital is flooding the robots. The teams solving data collection are the real shovels. cobalt-teleop.github.io x.com/Cointelegraph/status/2…

Imagine being able to move and teach a real robot just by opening a simple app on your phone. It’s kind of beautiful how $COBALT is turning the most ordinary object we carry every day "the smartphone" into a serious instrument for robot learning. No fancy hardware. No closed labs. Just people around the world using what’s already in their pockets to help build better robots. This is how @animesh_garg and the COBALT team actually building the future of AI robotics. cobalt-teleop.github.io

A robot watching someone pour coffee solves four problems at once: Time. Space. Physics. Embodiment. The body doesn’t match. The physics is implicit. Yet the signal is rich enough to learn from. If it pays off, the cost of teaching robots new skills falls through the floor.

Robotics is the next AI frontier. The winners won’t just be the companies building robots, but the platforms supplying the data they learn from. That’s why I’m paying attention to $COBALT and what @animesh_garg is building. cobalt-teleop.github.io/

For two centuries, making one of something cost far more than making a thousand. This constraint has driven almost everything that gets built at scale. robot labor breaks the assumption: small batch production starts to cost what a batch of a thousand does. What that does to where things get made, and who makes them? Read the blog on what labor as a utility could unlock? This is such a clear and powerful framing. Physical AGI doesn’t just automate tasks, it fundamentally rewires the economics of production by turning labor into something closer to a utility: scalable, location-independent, and no longer the dominant variable cost. $COBALT’s approach to crowdsourced, smartphone-based teleoperation is exactly the infrastructure that makes this vision actionable. By democratizing high-quality demonstration data at global scale (already proven with thousands of demos across countries in days), it supplies the “what” and “how” needed to train the robots that deliver this abundance without the old hardware and cost barriers. Excited to see how these ideas play out as the pieces come together. Brilliant post! cobalt-teleop.github.io

Spot on. That first step is the whole game. Generating reference data is the part that just doesn't scale with the teleop hardware we have today. And everything downstream inherits it. Your RL experts your distillation your real world deployment all of it is capped by the coverage you started with. So the real unlock is cheap diverse demos at scale. That's what makes COBALT worth watching. Crowdsourced teleop without the expensive rigs. Curious how far the loco-manipulation envelope stretches once data stops being the constraint. cobalt-teleop.github.io @animesh_garg $COBALT

“why humanoid” gets another case study? Expansion of capability envelope of humanoids and importantly in a very data and compute efficient manner. This is spot on. The future of capable humanoids will depend heavily on scalable, high-quality data collection, exactly what makes platforms like COBALT so valuable. Being able to gather diverse, real-world demonstrations efficiently (without expensive hardware) is going to be key to making these sim-to-real pipelines actually work at scale. Excited to see where this direction goes. $COBALT cobalt-teleop.github.io