Dyna-2.1: A Physical Agent for End-to-End Workflows
A robotics company has introduced Dyna-2.1, a full-stack physical agent built around a new semi-humanoid robot called Taku that can complete hour-long, non-linear manipulation workflows without human intervention. The system targets a shift in how businesses buy robotics: customers pay for a role, not a task, and the company said its commercial conversations show that customers want "a robot that can become a whole employee."
The announcement matters because stationary robot tasks, while production-ready, still require a person nearby to refill bins, clear stacks, and handle transitions between sub-steps. Dyna-2.1 is designed to close that gap by combining a wheeled, human-proportioned body with a three-layer learning system that handles control, physical skills, and workflow reasoning.
Meet Taku: built for reach, not just precision
Taku, named after the Japanese word Takumi meaning master craftsman, has a human-shaped body above the waist, a folding lower body for reaching low and high shelves, and two 7-degree-of-freedom arms on a stable four-wheeled base. The company sized Taku like an average person so that human motion-capture data - recorded wrists, elbows, and chest positions - maps naturally onto the robot's reachable poses.
The hardware choice reflects the workflow problem. A hotel laundry room has washers, dryers, a folding table, and shelves spread meters apart, with work happening at every height: leaning into a dryer drum, reaching into a washer for a leftover towel, crouching to a bottom shelf. The company calls this loco-dexterous manipulation - coordinated whole-body movement plus dexterous manipulation - and said it is difficult for table-mounted bimanual arms or elevator-style mobile manipulators to handle.
Why workflows are unforgiving
A laundry cycle chains about 79 steps, according to the company's analysis. At 95% reliability per step, a cycle almost never finishes without help. The math is stark: small failures compound, and some mistakes undo earlier work, like a dropped towel that must go back in the wash. The company chose laundry partly because most failures are recoverable, but recovery itself costs time and must still avoid unrecoverable errors.
The workflow is also non-linear. Towels move linearly from wash to dry to fold to stack, but the robot does not. Each machine finishes on its own schedule, and a finished machine left idle is lost capacity. Folding must be interruptible. The system breaks a cycle into thirteen decision points, several of which depend on information observed minutes or hours earlier - when the washer started, which load is in which machine, which shelf has room.
Three layers: controller, policy, and orchestrator
Dyna-2.1 divides responsibilities by timescale. A whole-body controller, trained with reinforcement learning in simulation, converts task-space target trajectories into joint targets and wheel velocities at 100 Hz. The DYNA-2 policy, an improved version of the company's earlier world-action model, turns the current step into whole-body target trajectories. A vision-language orchestrator tracks the workflow, decides the next step, and steers the policy to carry it out.
A key design choice is the Unified Robot Representation (URR), a data interface that describes a body by its wrist, elbow, chest, and footprint poses in a locally consistent coordinate frame. Both people and robots can produce these poses. The company pre-trained its action model on one million hours of human video mixed with robot fleet data, using whole-body poses tracked from video as targets. Human pre-training already showed results in earlier systems: DYNA-2, pre-trained on human video alone, passed 87% of customer acceptance tests zero-shot on a stationary robot at unseen sites, versus 46% for DYNA-1.
The controller and task form what the company calls a virtuous cycle: more demonstrations make a better controller, and a better controller records better demonstrations for the policy. The company also found that oscillations in controller response consistently harm policy learning, so "no oscillation" is a hard requirement in controller tuning.
Long-term memory and recovery
The orchestrator compresses visual progress into a text-based long-term memory: which steps are done, whether each washer and dryer door is open or closed, how many towels have been folded, which machine is running. This lets Taku pick up folding where it left off after attending to washer and dryer tasks. The orchestrator runs at a lower frequency than the policy, so it can use more context and reasoning time without slowing the physical execution layer.
Recovery is where the system's teachability pays off. The company said its stationary folding deployments now reach production bars in as little as three days, down from weeks or months of on-site engineering. Dyna-2.1 learns new skills with far less demonstration data than previous systems, with server servicing and drink retrieval cited as recent examples. The same teaching recipe carries across tasks with different demands on whole-body dexterity and precision.
Why this matters for operations and research professionals
For operations managers, the workflow framing changes the purchase decision. A robot that handles a single stationary task still needs a human babysitter for bin refills and stack clearing. A robot that completes hour-long workflows with infrequent, asynchronous handoff can be evaluated as a shift worker, not a tool. The company's metric - how long a workflow a model can handle without intervention - echoes the 50% task-completion time horizon proposed by METR for language agents, and it gives operations teams a concrete way to compare physical AI systems against labor costs.
For researchers and HR professionals tracking automation's impact on roles, the shift from task-level to workflow-level autonomy signals which jobs face nearer-term pressure. The company's deployment focus on commercial laundry and its plan to build a deployment flywheel - where every piece of site data improves the full system - suggests the gap between pilot and production is closing for physical AI in structured environments. Professionals evaluating AI agent courses or building skills for managing automated operations may find the workflow-reliability framing more useful than benchmark success rates. Those in operations leadership roles can explore AI for operations managers courses to understand how systems like Dyna-2.1 reshape staffing and process design decisions.
Your membership also unlocks:
