Dyna Robotics' new Taku humanoid is built for a job, not a set of tasks

HMHOctober 5, 2026

Dyna Robotics wants to sell businesses a robot employee, not a robot task. Taku is the hardware making that pitch real: a wheeled semi-humanoid that chains 79 subtasks across an hour of laundry work with no human standing by.

Load the washer. Wait for the cycle. Transfer wet towels to the dryer. Wait again. Fold. Stack. Shelve. One hour of hotel laundry work, done by a robot without a single human standing by. That's what Dyna Robotics is claiming with Taku, and it published the uncut footage to prove it.

The company unveiled Taku on September 29, 2026, alongside Dyna-2.1, the AI model that powers it. The robot is a semi-humanoid: human-proportioned upper body with 2 seven-DoF arms, a folding lower body for reaching high and low, and a base with 4 steerable wheels. It has 21 degrees of freedom total. Low-ratio planetary actuators replace conventional harmonic drives to support higher speeds and accelerations ... a deliberate engineering choice, not a spec sheet afterthought.

The name comes from Takumi, the Japanese word for master craftsman. Whether it lives up to that name in real deployments remains to be seen. But the claim is ambitious: Taku isn't supposed to do a task. It's supposed to do a job.

The math behind "fully autonomous"

Here's a number worth thinking about: to complete an hour-long laundry workflow — 79 subtasks, 13 decision points — without any human intervention, Taku needs a very high success rate: roughly 99.8% per step.

At 95% per-step, the probability of completing the whole cycle unassisted drops to about 1.7%. Dyna published that math openly.

The three-layer AI stack is designed specifically for this reliability challenge. A whole-body controller trained via reinforcement learning runs at 100 Hz, converting high-level targets into joint commands and wheel velocities. Above that sits the DYNA-2 world-action model, running at roughly 5 Hz, generating whole-body movement targets from perception. On top of both: a vision-language workflow orchestrator that tracks state, maintains memory across the full hour, and decides what comes next.

The orchestrator handles the 13 decision points — "is the washer done?", "which shelf is available?", "does this towel need refolding?" — without a human in the loop.

Cross-embodiment learning

The technical piece that makes Dyna's claims about deployment speed plausible is their Unified Robot Representation (URR) — a shared coordinate system describing robots and humans via wrist, elbow, chest, and footprint poses rather than joint angles.

Because it's body-agnostic, the same training signal works across human video, teleoperation data, and motion capture. That's why Dyna says new sites now reach production performance in 3 days, versus the weeks or months it took with earlier versions of the system.

The previous DYNA-1 and DYNA-2 systems were already deployed for stationary folding in hotels, restaurants, laundromats, and gyms. Taku extends that to mobile workflows: the robot can work at table height, crouch toward floor-level shelves, reach into washers, and transfer between equipment across a real room.

Who's behind it

Dyna Robotics was founded in 2024 by Lindon Gao, Jason Ma, and York Yang.

Gao and Yang previously built Caper AI, a grocery checkout robotics startup acquired by Instacart for $350 million in 2022. Ma came from Google DeepMind, where he worked on AI manipulation research.

The company raised a $23.5 million seed round in March 2025 led by CRV and First Round Capital. Six months later, in September 2025, it closed a $120 million Series A — led by Robostrategy, CRV, and First Round Capital, with NVIDIA's NVentures, the Amazon Industrial Innovation Fund, Salesforce Ventures, Samsung Next, and LG Technology Ventures all participating.

Total raised: roughly $144 million.
Post-Series A valuation: $600 million+.
The company has about 114 employees.

That's a strong investor list for a company barely 2 years old. NVIDIA and Amazon both have strategic reasons to bet on embodied AI platforms that could run on their hardware or inside their logistics networks.

The business model framing

Dyna's commercial pitch is worth noting: "Customers pay for a role, not a task." The robot takes over a whole workflow for a shift. Not "fold this towel" — "be the laundry room attendant."

That framing matters because it changes how buyers think about value. A task is a line item. A role is headcount. And in hotels and laundromats right now, those are real positions that are hard to staff and expensive to fill. If Dyna can reliably occupy one of those slots, the ROI calculation becomes straightforward.

"Our models continuously improve with each customer deployment, generating high-quality data," CEO Lindon Gao said in the Series A announcement. "We are observing true generalization as our robot enters new environments; it simply works out of the box, with no additional data."

That last claim — "works out of the box" in new environments — is the one to watch. Dyna's early deployments in hospitality suggest they're at least partway there. Whether it scales cleanly as the environments get messier is the real test.

The 3-day site onboarding figure suggests they're making progress on that front. Whether Taku can sustain those numbers at scale, across the full messiness of real commercial kitchens, hotel corridors, and laundry rooms, is the question 2027 will start to answer.