We had the realization earlier this year that the true value of general-purpose robots is in solving end-to-end workflows instead of isolated, stationary tasks. Much like LLM coding, autocomplete was cool, but it really took off when coding agent became capable enough to own an entire project end-to-end. Marching towards the vision that a robot should become a coworker that you can depend and trust for a whole shift, we have revamped our entire stack, built our friendly and cute semi-humanoid Taku, and designed an agentic system that seamlessly reacts, executes, and plan. Dyna-2.1 combines a very capable hardware with full-stack modeling to deliver a physical agent that learns, adapts, recovers extremely efficiently. I hope this becomes an inflection point for the whole industry, as we shift away from demos and towards building a product that can truly increase physical productivity in the world. General-purpose robots are coming, and i'm excited for what to come next

Mobility is the key unlock. Tabletop tasks, no matter how reliable, only go so far if the robot needs to be fed and babysat. Mobility completes the autonomy loop. @DynaRobotics is taking their destiny into their own hands by designing their own semi-humanoid. Also cool to see the engineers talking about their work in this video.

From simulation to real-world autonomy. 🤖 Dyna Robotics trained Dyna-2.1 in part with NVIDIA Isaac Lab, where thousands of simulated robots learned whole-body control in parallel. Learn how Dyna-2.1 handles an entire hotel laundry workflow from start to finish.

Check out the deep dive on Dyna-2.1 from @kun_h____ !

Here’s a more personal breakdown of how we built Dyna-2.1 - the first Physical Agent to achieve 1-hour-long, dexterous whole-body autonomy. A lot of the system came down to a few fairly fundamental design choices.

Taku in dark neon light server room is such a vibe

This is the direction robotics needs to take if we want it to fulfill its potential. It is not about humans carefully setting up a perfect environment so the robot can just perform specific tasks. It needs to handle the entire end to end workflow as if you have hired a human employee. That is true automation and the right way to use robots. The next step I want to see is the physical agent improving on the job: tighter traversal, better task sequencing, stronger operational judgment over a shift. Congrats to the team on the launch. Looking forward to seeing this in real environments!

Dyna just unveiled Dyna-2.1 and its new robot, Taku, bringing its model and hardware together as a Physical Agent designed around customer workflows. They’ve also opted for a wheeled semi-humanoid design. Taku has a human-shaped upper body, a folding lower body for reaching high and low, an extremely stable base with four steerable wheels, and two 7-DoF arms. Its work is mostly precise reaching and manipulation, not legged movement. That’s the practical reason so many robots use wheels. Dyna-2.1 coordinates robot control, manipulation skills and workflow planning to handle real-world jobs autonomously from start to finish. In the company’s one-hour, uncut hotel laundry demo, Taku loads and starts the washer, transfers towels to the dryer, then folds, stacks and shelves them.

Presenting research from @berkeley_ai Humanoid Intelligence Center. DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation.🖐️🤖 Dexterous manipulation requires touch, yet multi-finger tactile data remain scarce and expensive to collect at scale. Through continual vision-to-touch learning, DexTacWAM adapts a pretrained video world model into a visuo-tactile world model for predictive multi-finger contact modeling and action generation. Across six contact-rich dexterous manipulation tasks, DexTacWAM achieves the highest score on every task, averaging 70.6 vs. 38.0 for the strongest baseline. 🔬 Why not simply inject tactile features into the action policy? With the same multi-finger tactile encoder, policy architecture, and training procedure, replacing the tactile world-model latent with direct tactile features drops the 4-task mean from 74.7 to 26.6. Takeaway: the benefit does not come simply from providing tactile observations to the policy, but from making multi-finger contact evolution part of the predicted world state. All code, model weights, and datasets are now open-sourced! 🌐 Project: dextacwam.github.io/ 📄 Paper: arxiv.org/abs/2609.24976 💻 Code: github.com/dextacwam/DexTacW… 🤗 Weights&Data: huggingface.co/collections/J… Author List: @Jensen_Yuan, @zekaiw04, @sbn_epiphany Haoran Lu, @trevordarrell, @Ismini_L, @Wei_ZHAN_

My best robot demo of the year. From the bare-bones 7-DoF frame and wobbly AGV I saw last year to this smooth, compliant, full-size humanoid robot, what an incredible leap. Congrats to Dyna folks! Can’t wait to see what’s next

Dyna just announced two things at once: a new brain and a new body. The brain: Dyna-2.1, which they call the world's first "physical agent" that runs an entire workflow on its own, not just a task. Customers don't buy tasks, they buy a whole AI-employee. A robot that folds for hours but needs a human to refill and reset it is just a "robot babysitter." The body: Taku, a semi-humanoid with a wheeled base, human-shaped upper body, and folding torso, so it can grab a dropped towel off the floor, reach into a washer, and stack high and low shelves in one flow. Under the hood: RL whole-body control (100Hz), the DYNA-2 policy learned from human video, and a VLM orchestrator that reasons, remembers, and picks the next step. The new bar isn't what a robot does in a clip. It's how long it does the job without intervention. The demo: an uncut 1-hour hotel laundry shift, fully autonomous:

The quality of the execution and team collaboration has been amazing. I can’t imagine how much we can achieve in the coming months. A truly incredible stage to be at and more to come for sure!

not just about laundry

Nice attention to details :) In real deployments, it’s these little details that make or break a workflow

this is the kind of acceleration my laundry pile desperately needs lmao the cool part is their agent remembers where it left off, switches between machines and reasons out what to do when something goes wrong they literally pulled its hand off the washer and messed with the settings mid-task but it went back and set the correct cycle anyway also one of the cuter humanoids i've seen. i think robots are going to get the same cute treatment as agents like dots and muse.

This release is pretty cool for anyone who cares about robotics: • It comes with a vision-language planner • This planner chooses tasks and remembers progress • It translates tasks into whole-body motion • An RL controller executes movements at 100 Hz • A shared motion representation enables learning from humans • Error recovery keeps long workflows moving An hour-long autonomous laundry demo. I think this is the longest I've seen a robot stay focused on a task.

We are entering the agentic era of robotics

We had a small private viewing of Taku last week with friends, colleagues, and mentors.. Everyone’s first reaction seeing Taku work in person was “magical”. You also get to take cute pics with Taku, you don’t want to miss the next one

One of the best parts of working in deployments is seeing customer feedback shape what we build. Our stationary robots can do a lot, but some of the jobs our customers want help with are quite literally out of reach. Those conversations have helped shape our new hardware, including Taku, so we can take on more of the everyday work our customers need done. Watching @kun_h____ , @ethanwjahn , and the whole team take Taku from “nothing works” to an incredibly reliable robot in just a few months has been one of the most impressive feats of execution i've seen 🐐🐐. Excited to share what Taku can do with DYNA 2.1.

I started working on this project at the beginning of 2026, when our first Taku prototype arrived at our US office. I've been thinking about what to write for this release: the breakthroughs, the design choices, the research contributions that might interest the broader robotics community. But after nine months of building a system far more complicated than stationary manipulation, what has stayed with me most is how much more humans can do as a team with strong alignment (the decision to use URR as the main interface across the stack being one example) and a shared direction (automating an hour-long laundry workflow) than as individuals. DYNA 2.1 is a significant milestone for Dyna. Literally every person here helped make it possible, across data ops, research, hardware, ML infra, and everything in between. Special thanks to my project co-lead @ethanwjahn, the most full-stack person I know. (You can still find all the interesting technical stuff in our tech blog post: dyna.co/dyna-2.1)