We are releasing Dyna-2.1, the first Physical Agent that achieves reliable super long-horizon whole-body autonomy. It combines our brand-new semi-humanoid hardware with an agentic system built around Dyna-2 to handle ultra-long real-world workflows.
Here is an uncut footage of Dyna-2.1 completing an entire hour-long laundry room workflow, just like a human does.
Sep 29, 2026 · 4:24 PM UTC
We have previously demonstrated commercial-grade mastery on individual *tasks*, but with Dyna-2.1, our robots are becoming agentic whole employees that complete workflows wherever they are deployed. Customers pay for complete roles, not isolated tasks.
Real-world workflows such as laundry-room operations are extremely complex. Even with 99% success rate on each individual step, the success rate for a whole laundry cycle is below 50%. Workflows also require chaining non-linear sequences of tasks, which is far beyond what a single action model (e.g., Dyna-2 WAM) can and should do.
To meet the challenges of workflow automation, we optimized and innovated across the full stack from designing our own hardware, controller, to building highly efficient models for action and reasoning traces.
First, introducing our new robot, Taku. It is inspired by takumi (匠), the Japanese tradition of mastery through craft. We designed and built Taku from first principles to optimize for data transferability and robot learning.
Using large-scale GPU parallelized reinforcement learning in Isaac Sim, we are able to obtain a robust whole-body controller for Taku optimized for precise loco-manipulation.
Taku is designed to be able to transfer physical skills from human videos. Combined with Dyna-2 WAM that’s pre-trained on millions of hours of human videos, Taku is a fast learner for new tasks.
With just 30 minutes to a couple hours of on-robot data, Taku can complete new whole-body tasks that require vastly different motion primitves.
Dyna-2.1 makes significant strides in making general-purpose robots useful in the real world. Our next milestone is to take this brand new system to real-world customer sites and establish a deployment flywheel, in which every piece of deployment data improves the full system. Our long term vision is to build a self-improving physical agent. And Dyna-2.1 is just the beginning.
Read our full technical report for deep dives and additional details: dyna.co/dyna-2.1
Long workflows are the true test for a robot. One that can get through a whole job without a reset is what a lot of operations leaders have been waiting for. Come on my show to talk about it: linktr.ee/evankirstel
121
