Dyna Robotics

A proud moment for the team behind Taku! 💙 Thank you RoboStrategy Advisors for bringing us together at Nasdaq.

AI runs on compute. Compute needs scaled infrastructure. We are teaming up with Lambda to give AI infrastructure some helping hands.

I've watched a lot of robot demos this year. Turns out I was grading them on the wrong number. Just a few days before announcing Dyna 2.1 and their semi-humanoid Taku, Lindon Gao joined me on the Dirty Jobs main stage and teased for our invite-only audience: — Why Speed, Throughput, Autonomy are most critical to what they measures now — Why hardware remains the bottleneck in physical AI today and the semi-humanoid opportunity — Why there's no such thing as bad data, only lazy labeling — Why deployment data is essential to building the world's best robotics model Lindon shared how their robot at Dyna folded 850 napkins a day at a 99.4% success rate, and when he took it to customers none of that mattered to them. They wanted throughput, quality and autonomy, and to work with a team who understood how those three trade off against each other. I loved his framing of the robot babysitter. Look at almost any demo and there is a person just out of frame swapping the bins every half hour. You can hide that in a well edited launch video but nobody is going to pay for that. BTW Lindon built Caper and sold it to Instacart for $350M, so he has built for scale before. Together, we got into why he still owns pre-training, why they started with small businesses where a mistake costs you a towel vs. something more serious, and why lack of reliable hardware is what's holding Physical AI back right now. 99% is not the finish line. It's the starting line. Episode 103 of CLIMB is out. Link in the comments.

Launching Dyna-2.1: the first physical agent that runs an entire workflow on its own. Not a single task. The whole job. Soon, you'll be able to onboard a robot the way you'd onboard a new hire, and they will take the job from there. We put it to work on a real hotel laundry shift. 24/7, end-to-end, no human interventions. Dyna-2.1 unloads the dryer, loads and starts the washer, folds towels, stacks them by size, shelves them, and starts the next load. Laundry is the starting line. Next, Dyna-2.1 will scale to hundreds of customer sites, and every hour it works will make the whole system better. Read our full technical report for deep dives and additional details: dyna.co/dyna-2.1

Meet your next new hire. Hits the ground running on day one. Works 24/7. Doesn't need anyone looking over its shoulder. Stay tuned 👀

After so many demos, models, pilots, robots are still struggling to land real deployments with real customers. Until now. Today we're sharing the behind the scenes of something a year in the making: Din Tai Fung, one of the highest revenue per location restaurant chains in the country, is rolling out Dyna robots across its restaurant network. This adds to our rollouts across hotels, logistics, data centers, and many other use cases, bringing our fleet to hundreds of robots by the first half of 2027. As far as we know, this is the first time a robot foundation model company has taken a manipulation pilot all the way to scaled deployment. We crossed the ROI threshold. And we're just getting started. It's easy to show a demo. What comes after are the edge cases, recovery, uptime, and the small failures that only show up once a robot has done the job every day for months. In the field. We spent the past year in that "after" with Din Tai Fung, running their napkin operations day in and day out. We moved the eval into production and graded every episode, at every site, on the customer's definition of good, over a terabyte of data a day. A new deployment now goes from setup to production ROI in as little as 3 days. It's been a wild year. And we're convinced that deployment is not pulling research away from the frontier, but rather where the frontier is getting built. Full blog post in the comments.

Today we're publishing how we trained Dyna-2 on over 1,000,000 hours of egocentric video, repeatably. Training a robotics foundation model at this scale is unprecedented, and it forced us to rebuild our data and training infrastructure end to end. Most of what worked at ten thousand hours did not hold up at a million. • Ingestion throughput: our pipeline was capped at 14,000 episode-hours per week. At that rate, a million hours would have taken over a year. We rewrote it around DAG decomposition, staggered starts, and bin-packed batches, scaling throughput to 440,000 episode-hours per week, a 31x improvement. • Manifest construction: every training run depends on a manifest built from the full corpus. At a million hours, this took 48 hours before a run could even start. Moving to a single warehouse query cut that to under a minute, and load time dropped from 737 seconds to 12.4 seconds through memory-mapped columnar tables. • Data delivery: our training clusters don't always sit next to our data, and reading a petabyte-scale corpus straight from cloud storage exposes GPUs to egress latency and packet loss. An on-cluster NVMe caching layer cut the time to read one petabyte from 57.9 days to 5.8 days, sustaining ~2 GB/s per node and keeping GPUs at 98% utilization. • Training efficiency: we rebuilt our Muon optimizer to shard state inside each node rather than across the job, avoiding the slower inter-node fabric as node count grows. This made the optimizer step roughly 3x faster at scale. • Job resilience: preflight health checks and auto-restart from the last checkpoint mean the run survives failures instead of stalling on them, and new clusters can now be stood up in days instead of weeks. Million-hour training is not a single scaling problem. It's an end-to-end system redesign, and the infrastructure is what makes this scale repeatable rather than a one-off. Link to the full technical report: https://lnkd.in/gWxXK45X

Today we are introducing Dyna-2, a world-action model pre-trained on over 1,000,000 hours of human video. At this scale, for the first time, we uncovered several surprising scaling laws: - Human manipulation obeys a power scaling law across three orders of magnitude, from 1,000 to 1,000,000 hours, with no plateau in sight. - The most important result: the human scaling law implied a scaling law on robot data the model had never seen. Scaling human pre-training predictably improved zero-shot performance on robot data, and carried through post-training to on-robot evaluation - World modeling on large amounts of video data is what unlocks this cross-embodiment transfer, and choosing the right world-modeling objective leads to even better generalization with more data. Beyond the scaling laws, Dyna-2 shows new capabilities that exceed our prior models across the board. - Language following: it is the first model to clear our internal language-following benchmark. - Production-ready performance: Dyna-2 has exceptional robustness and zero-shot environment generalization. It continues to work under various disturbances. The same task attains 87% quality and throughput grades even when deployed in an entirely new environment. We believe Dyna-2 presents a new frontier in scaling embodied AI and charts a clear path toward achieving general embodied intelligence. Link to the full technical report in the comments.

A new breakthrough is coming. Stay tuned.