# Dyna-2.1

> Source: https://aiwiki.ai/wiki/dyna_2_1
> Updated: 2026-09-30
> Fact-checked: 2026-09-30
> Categories: Embodied AI, Humanoid Robots, Physical AI, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Dyna-2.1." aiwiki.ai, 30 Sept 2026. https://aiwiki.ai/wiki/dyna_2_1
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Dyna-2.1** (styled **DYNA 2.1** in the company's press release) is a robot system announced by [Dyna Robotics](https://aiwiki.ai/wiki/dyna_robotics) on 29 September 2026. It pairs a new wheeled semi-humanoid robot named **Taku** with a three-layer software stack: a reinforcement-learning whole-body controller, an improved version of the company's DYNA-2 [world action model](https://aiwiki.ai/wiki/world_action_model), and a vision-language "workflow orchestrator" that decides what to do next and keeps a running text memory. Dyna calls the combination a "physical agent" and presents it as a step from automating single tasks, such as napkin folding, to handing a robot a whole job for a shift.[1][5] The launch demonstration was what Dyna calls an uncut, roughly one-hour run in which Taku worked through a hotel laundry room on its own: loading and starting washers and dryers, unloading them, folding towels, and shelving the stacks.[1][2]

Dyna describes Dyna-2.1 as "the first Physical Agent that achieves reliable super long-horizon whole-body autonomy" and, in its research post, as, "to the best of our knowledge," the first physical agent to perform hour-long, non-linear loco-dexterous manipulation workflows. These are the company's own claims.[1][2] Part of the controller training ran in [NVIDIA Isaac Sim](https://aiwiki.ai/wiki/nvidia_isaac_sim), and NVIDIA's robotics account promoted the launch on that basis.[1][4]

## Overview

| Item | Detail |
| --- | --- |
| Developer | [Dyna Robotics](https://aiwiki.ai/wiki/dyna_robotics) (Redwood City, California)[5] |
| Announced | 29 September 2026 (research post, press release and X thread); the post is listed as 09.28.26 on Dyna's research index[1][2][5][9] |
| What it is | A "physical agent": the Taku robot plus a whole-body controller, an improved DYNA-2 policy and a vision-language workflow orchestrator[1] |
| Robot | Taku, a semi-humanoid with a human-shaped upper body, folding lower body, four steerable wheels and two 7-degree-of-freedom arms[1] |
| Headline demo | A roughly one-hour autonomous hotel laundry-room workflow, which Dyna calls uncut; the launch video shows it at 1x to 10x speed[1][2][10] |
| Other demonstrated skills | Server servicing and retrieving a drink from a refrigerator[1] |
| Simulation | Whole-body controller trained with reinforcement learning in NVIDIA Isaac Sim[1][4] |
| Availability | No price announced; launch materials describe deployment status differently (see Deployment status)[1][5] |

## Background: from tasks to workflows

Dyna describes its earlier deployments as stationary tasks: its DYNA-1 and DYNA-2 models ran stationary folding stations in commercial deployments, including napkin folding for restaurants and commercial laundry.[1][8] The Dyna-2.1 post argues that such stations still need a "robot babysitter". Its example is DYNA-1's 24-hour napkin-folding run, which had no interventions on the folding task itself, while a person still had to refill the napkin bin and clear finished stacks.[1] From customer conversations, Dyna says it concluded that "customers pay for a role not a task", meaning a robot that runs a whole workflow for a shift without a second person standing by.[1]

Dyna defines a workflow as "a self-contained business process that allows infrequent, asynchronous handoff": a person hands it over, leaves, and checks back hours later.[1] The post lists three capabilities it says workflows require: a workspace as large as the human role being filled, "teachability" (learning new skills quickly and making them reliable), and reasoning, because workflows branch and depend on conditions.[1] It argues that meeting all three at once is no longer feasible for "a single action model deployed on a limiting hardware platform", and that a full stack is needed, from hardware design to policy learning.[1]

The reliability argument is multiplicative. Dyna says a laundry cycle chains about 79 steps, and that at 95% reliability per step a cycle "almost never finishes without help".[1] (At 95% per step, 0.95 to the 79th power is under 2%.) Its X thread gives a second version: even at 99% per step, a whole laundry cycle succeeds less than half the time.[3] Dyna's conclusion is that recovery from errors has to be taught step by step, not assumed away.[1]

## Taku hardware

Taku's name comes from the Japanese word *takumi* (匠), which Dyna glosses as "master craftsman" and, in its X thread, as "the Japanese tradition of mastery through craft".[1][3] ABC7 San Francisco, which got what it called an exclusive preview, reported the same derivation.[6]

| Feature | Description (as stated by Dyna) |
| --- | --- |
| Form factor | Semi-humanoid: a human-shaped body above the waist on a wheeled base[1][5] |
| Lower body | Folds so the robot can reach low and high[1] |
| Mobility | "An incredibly stable base on four steerable wheels"; the press release says the wheels let it move "without risk of toppling over"[1][5] |
| Arms | Two 7-degree-of-freedom arms[1] |
| End effectors | Arms compatible with parallel-jaw grippers or dexterous hands (press release)[5] |
| Actuators | Low-ratio planetary actuators in the arms, chosen for higher speed and acceleration than the [harmonic drives](https://aiwiki.ai/wiki/harmonic_drive) of conventional arms[1] |
| Sizing | Above the wheels, sized like an average person, so tracked human wrist, elbow and chest positions land near poses Taku can reach[1] |

Dyna says it chose wheels over legs because most steps in its target workflows need reach and precise manipulation rather than walking.[1] In the laundry room, washers, dryers, a folding table and shelves stand meters apart, and the work runs from reaching an arm's length into a dryer drum to crouching at the bottom shelf or reaching above head height. Dyna argues this is hard for "a table-mounted bimanual arm or an elevator-style mobile manipulator".[1] The company's X thread says Taku was designed "from first principles to optimize for data transferability and robot learning".[3]

Dyna did not publish Taku's height, weight, payload, battery life or price in the research post or press release.[1][5] An early prototype of Dyna's semi-humanoid had appeared in the August 2026 DYNA-2 post, where it ran a language-following "targeted drink retrieval" task.[7]

## The three-layer stack

Dyna splits control of Taku into three models, divided "by the timescale of each decision and by how each capability can be learned most efficiently".[1]

| Layer | Role | How it is trained |
| --- | --- | --- |
| Whole-body controller | Turns task-space targets for the wrists, elbows, chest and footprint into joint targets and wheel velocities at 100 Hz | [Reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) in simulation (NVIDIA Isaac Sim), seeded with human motion data[1] |
| DYNA-2 policy | An improved version of Dyna's DYNA-2 world-action model; turns the current step into whole-body target trajectories for the controller | Pre-trained on one million hours of human video mixed with robot fleet data, then fine-tuned on robot data[1] |
| Workflow orchestrator | A [vision-language model](https://aiwiki.ai/wiki/vision_language_model) that tracks the workflow, decides the next step and steers the policy; keeps a long-term memory as text | Built on a VLM pre-trained on text and images, post-trained on the workflow's branches[1] |

The press release describes the same design as "a vision-language orchestrator that reasons through complex workflows and a whole-body controller that coordinates driving, reaching, bending, and lifting", underpinned by DYNA-2, which it says was "trained on a million hours of human and robot data".[5]

### Unified Robot Representation

The interface between the layers is what Dyna calls the **Unified Robot Representation (URR)**. URR describes a body by its wrist, elbow, chest and footprint poses in a locally consistent coordinate frame; a sequence of these poses describes motion.[1] Dyna lists three properties:[1]

- It is embodiment-agnostic. The same format covers a human, a humanoid, a semi-humanoid or a tabletop bimanual robot, using more or fewer body parts depending on what is present or observed.
- It captures the task-space constraints Dyna considers most important for coordinated manipulation: hand placement, arm configuration, torso posture and where the body stands.
- It is lossless across equivalent action formats. The poses can be recovered from formats such as relative-pose actions, as long as the reference frame and starting pose are kept.

Because people and robots can both produce URR poses, Dyna converts [teleoperation](https://aiwiki.ai/wiki/teleoperation) data, egocentric human recordings and [Universal Manipulation Interface](https://aiwiki.ai/wiki/universal_manipulation_interface) (UMI) data into URR for training. At inference, the model's actions are decoded back into URR target trajectories for the controller.[1] Dyna contrasts this with most whole-body teleoperation stacks, which fit human motion to robot joints before a tracking controller runs. It says that joint fitting can pull a hand away from the object it touched, so the video and the action labels disagree; URR keeps the demonstrated motion in task space and leaves joint coordination to the controller.[1] The same interface lets a person and the model hand control back and forth, which Dyna uses to record corrections.[1]

### Whole-body controller

The controller learns balance, posture and motion tracking with reinforcement learning in simulation. The post's Video 3.3 shows "thousands of simulated Taku robots training in parallel in NVIDIA Isaac Sim", and Dyna's X thread describes "large-scale GPU parallelized reinforcement learning in Isaac Sim".[1][3] Instead of relying mainly on exploration and hand-tuned rewards, Dyna feeds its URR recordings to the controller as pose targets, alongside synthetic targets, to give it human pose priors. Every new demonstration therefore also enlarges the controller's training set. Dyna reports that the growing human prior sped up convergence and improved the controller in offline metrics and teleoperation tests, but gives no figures.[1]

Dyna's other design points for the controller:[1]

- It tracks elbows and chest as well as hands, so the forearm stays clear of shelf edges and machine doors.
- Unlike inverse kinematics, it is dynamics-aware, tolerates infeasible commands, and encodes priorities (such as hand accuracy over chest accuracy) as reward weights.
- Link masses, joint friction and damping, and actuator gains and delays are randomized in simulation to narrow the gap to real hardware.
- It is tuned to be overdamped. Dyna treats "no oscillation" as a hard requirement because it found oscillation consistently harmful to policy learning. It cites the April 2026 paper *Tune to Learn* by Antonia Bronars, Younghyo Park and Pulkit Agrawal, which found that behavior cloning benefits from overdamped gains.[1][12]

Dyna frames the controller and the policy as a loop: more demonstrations make a better controller, and because every demonstration is recorded through the controller, a better controller produces cleaner motion for the policy to learn from.[1] See [whole-body control](https://aiwiki.ai/wiki/whole_body_control) for the wider research area, which Dyna notes has mostly targeted legged locomotion with imprecise arms.[1]

### DYNA-2 policy and teachability

Dyna defines teachability as "how cheaply a new skill, or a new customer's procedure, reaches production reliability".[1] The middle layer is an improved DYNA-2 policy, pre-trained on one million hours of human video mixed with robot data from Dyna's fleet, using whole-body poses tracked from the video in URR as targets.[1] Because those targets are already in the format the policy outputs on the robot, Dyna says fine-tuning on robot data only has to teach the policy Taku's specific body, and the human video shows the policy towels, machines and rooms that its robot data never covered.[1] Its X thread says that with "just 30 minutes to a couple hours of on-robot data", Taku can complete new whole-body tasks that need very different motion primitives.[3] The post shows server servicing and drink retrieval as examples of recently learned skills, but gives no data counts or success rates for them.[1]

Two teachability figures in the post come from Dyna's earlier, stationary systems, not from Taku:

- **87% versus 46%.** DYNA-2 passed 87% of customer acceptance tests zero-shot at sites it had seen no data from, against 46% for DYNA-1.[1] The DYNA-2 post gives the method: pass rates from on-site reporting graded against customer acceptance criteria by operators not involved in model development, with both models post-trained on the same task data for the same number of steps.[7]
- **Three days.** In Dyna's stationary folding deployments, a new site "now reaches its production bar in as little as three days, down from weeks to months of on-site engineering".[1] Dyna's August 2026 deployment post described three days as the shortest time from setup to meeting a customer's production ROI bar.[8]

To keep older robot data useful across hardware changes, Dyna leaves joint-level motion and limits to the controller. A hardware revision then means retraining the controller in simulation, "with no robot time", instead of changing the policy's action space.[1] Because each demonstration still carries the dynamics of the controller that recorded it, Dyna adds a metadata prompt telling the policy which controller produced each demonstration. It builds this on DYNA-2's language steerability, so a single checkpoint can be steered to the controller used in deployment. It also conditions the model on future commands during asynchronous inference, which gives implicit information about controller dynamics.[1]

### Workflow orchestrator and memory

The top layer is a vision-language model, pre-trained on diverse text and images and post-trained on the laundry workflow's branches.[1] It runs at a lower frequency than the policy, which keeps executing the current step. Dyna says this means a larger, slower vision-language model could be swapped in without changing the layers below, and that each decision can use more context and more time to reason.[1] People can give the orchestrator short instructions, which it expands into steps the policy can carry out. Dyna says this means some changes to a customer's procedure need no new demonstrations.[1]

At each branch point (which machine to load or unload, or whether a stack has enough towels to shelve), the orchestrator picks the next step and steers the policy. When a completion check fails, it decides whether to repeat the step, try another way, or change the plan; the policy carries out the recovery.[1] Because no training dataset records what has happened in a particular room that day, the orchestrator keeps a long-term memory as text: which steps are done, whether each washer and dryer door is open or closed, how many towels have been folded, and which machine is running. Dyna's example is that Taku resumes folding at the right towel count after breaking off for washer and dryer tasks.[1] In a video embedded in the post, Dyna staff Ling Li, Ming Qin and Chet Bhateja discuss the design of the reasoning and language-following system.[1]

## The laundry demonstration

Dyna uses a hotel laundry room as the running example. It says commercial laundry is one of the deployment verticals it has focused on, and that it chose laundry partly because most failures there are recoverable.[1] The post describes two problems a single task does not pose:[1]

1. **Long chains of subtasks.** Over an hour small things go wrong repeatedly: a grasp catches two towels, a corner folds under, a stack leans. Some errors undo earlier work; a dropped towel has to go back in the wash.
2. **Non-linearity.** Each machine finishes on its own schedule, and an idle finished machine is lost capacity, so folding has to be interruptible. Dyna breaks one cycle into **thirteen decision points**, several of which depend on information the robot saw minutes or hours earlier: when the washer started, which load is in which machine, and which shelf has room.

According to the press release, the demonstrated shift included loading the washer and dryer and turning the machines on; bending to reach the back of a dryer and unload bulk towels into a basket; snapping towels flat, folding them, and sorting them into stacks by size; crouching or reaching up to shelve finished stacks; and correcting physical errors, such as picking up a dropped towel from the floor or abandoning a bad grasp, "without requesting human help".[5] Separate clips in the research post show Taku finishing a washer cycle setting after a person perturbs the dials and pulls its gripper off the buttons, recovering from a double-towel grasp, and pausing folding to unload a dryer.[1]

The launch video Dyna posted on X and YouTube runs about 10 minutes 44 seconds.[2][10] On-screen labels mark the footage as autonomous and play it at speeds from 1x to 10x, with an elapsed-time counter that ends at about 1:00:47. Text overlays show the orchestrator's reasoning, for example "Both machines are running. Resume folding while the cycles run."[2][10] Dyna calls the footage uncut.[1][2] It did not say in the post or press release how many full-workflow runs it attempted or how many succeeded.[1][5]

## Other demonstrated skills

Beyond laundry, the research post shows Taku learning to service a server and to retrieve a drink, and a gallery of skills including pressing a button, placing a stack on a high shelf, inserting a server and picking a laundry pod.[1] The press release names food preparation, data center management and commercial laundry as examples of end-to-end workflows the system can manage.[5] ABC7 reported, citing the company, that the robot can load washers, fold towels, chop vegetables and stock shelves.[6]

## Evaluation: mean time between interventions

Dyna argues that per-episode success on benchmark tasks stops being informative once models work, because it "says nothing about whether the same model can be left alone with a shift". It proposes measuring how long a workflow a model can handle without intervention.[1] The post draws an analogy to [METR](https://aiwiki.ai/wiki/metr)'s March 2025 paper *Measuring AI Ability to Complete Long Tasks* (later retitled *Measuring AI Ability to Complete Long Software Tasks*), which proposed the 50% [task-completion time horizon](https://aiwiki.ai/wiki/metr_time_horizon) as a measure for language agents.[1][11]

The press release names the target metric as Mean Time Between Interventions (MTBI), "how long the machine operates before a human must step in to assist".[5] It says the system captures execution data as it runs, "from the reasoner's decision traces to the controller's joint loads", and feeds it back to raise the minutes of work per intervention.[5] Co-founder Jason Ma called MTBI "a far more useful benchmark than per-episode success rates on isolated tasks, which is widely used today but fails to reflect true commercial viability".[5] Neither the press release nor the research post reports an MTBI figure for Dyna-2.1.[1][5]

## Deployment status

The sources differ on deployment. The research post and Dyna's X thread both say: "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."[1][3] The press release, issued the same day, says "DYNA 2.1 is being deployed in hotels, laundromats, and restaurants", and ABC7 reported the company as saying the technology "is already being used in hotels, restaurants and laundromats".[5][6] Dyna's earlier stationary robots were already at customer sites: the press release says Dyna's robots "have been deployed at customers across multiple industries", and in August 2026 Dyna said the Din Tai Fung restaurant chain was rolling them out across its network.[5][8] The launch materials cited here do not name a customer site where Taku itself is operating. Dyna's stated long-term aim is a self-improving physical agent that absorbs diverse data and improves from its own experience.[1]

In the press release, CEO Lindon Gao said the goal is to make general-purpose robots commercially viable and that "Our customers need robots that complete an entire workflow for a full shift".[5]

## NVIDIA Isaac Sim

Dyna is backed by NVIDIA's venture arm, NVentures.[13] On 29 September 2026 the NVIDIA Robotics account on X wrote that "Dyna Robotics trained Dyna-2.1 in part with NVIDIA Isaac Sim, where thousands of simulated robots learned whole-body control in parallel", quoting Dyna's launch post.[4] Dyna's own account also credits Isaac Sim for the controller training.[1][3]

## Reception

Coverage on launch day was mostly the press release, syndicated through PR Newswire and outlets that republish it.[5] ABC7 San Francisco (KGO-TV) ran a segment on 29 September 2026 after what it called an exclusive preview of the robot, quoting Gao that Dyna is "bringing robots into real life" and "making it commercially viable".[6] Dyna's launch post on X had been viewed more than 400,000 times by 30 September 2026, and NVIDIA Robotics reposted it with its own commentary.[2][4]

The main figures in the launch (one million hours of pre-training video, the 87% acceptance rate, three days to a production bar, 30 minutes to a couple of hours of robot data for a new task) are all self-reported, and the acceptance and three-day figures refer to earlier stationary systems.[1][3][7][8]

## See also

- [Dyna Robotics](https://aiwiki.ai/wiki/dyna_robotics)
- [World action model](https://aiwiki.ai/wiki/world_action_model)
- [Robot foundation model](https://aiwiki.ai/wiki/robot_foundation_model)
- [Humanoid robot](https://aiwiki.ai/wiki/humanoid_robot)
- [Embodied AI](https://aiwiki.ai/wiki/embodied_ai)
- [Physical AI](https://aiwiki.ai/wiki/physical_ai)
- [Imitation learning](https://aiwiki.ai/wiki/imitation_learning)
- [Robot manipulation](https://aiwiki.ai/wiki/robot_manipulation)

## References

1. Dyna Robotics. "Dyna-2.1: A Physical Agent for End-to-End Workflows." DYNA Research, 29 September 2026. https://www.dyna.co/dyna-2.1
2. Dyna Robotics (@DynaRobotics). "We are releasing Dyna-2.1, the first Physical Agent that achieves reliable super long-horizon whole-body autonomy..." X, 29 September 2026. https://x.com/DynaRobotics/status/2104970523726033387
3. Dyna Robotics (@DynaRobotics). Dyna-2.1 launch thread on X (posts on step reliability, Taku, Isaac Sim, on-robot data and next milestone), 29 September 2026. https://x.com/DynaRobotics/status/2104970527903461782 ; https://x.com/DynaRobotics/status/2104970531435061712 ; https://x.com/DynaRobotics/status/2104970533263831072 ; https://x.com/DynaRobotics/status/2104970536015343791 ; https://x.com/DynaRobotics/status/2104971005865427434
4. NVIDIA Robotics (@NVIDIARobotics). "From simulation to real-world autonomy. Dyna Robotics trained Dyna-2.1 in part with NVIDIA Isaac Sim..." X, 29 September 2026. https://x.com/NVIDIARobotics/status/2105081108719440013
5. DYNA Robotics. "DYNA Robotics Launches DYNA 2.1 Physical Agent, a Semi-humanoid Robot that Completes Full Workflows such as a Commercial Laundry Shift." PR Newswire, 29 September 2026. https://www.prnewswire.com/news-releases/dyna-robotics-launches-dyna-2-1-physical-agent-a-semi-humanoid-robot-that-completes-full-workflows-such-as-a-commercial-laundry-shift-302892411.html
6. "Meet Taku: Bay Area company's AI robot loads washers, folds towels and stocks shelves." ABC7 San Francisco (KGO-TV), 29 September 2026. https://abc7news.com/post/meet-taku-redwood-citys-dyna-robotics-ai-robot-loads-washers-folds-towels-stocks-shelves/19888144/
7. Dyna Robotics. "Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models." DYNA Research, August 2026. https://www.dyna.co/dyna-2
8. Dyna Robotics. "Not Just a Model, But a Product." DYNA Research, August 2026. https://www.dyna.co/research/scaling-customer-deployments
9. Dyna Robotics. "Research" (index of research posts). https://www.dyna.co/research
10. Dyna Robotics. "DYNA 2.1: A Physical Agent for End-to-End Workflows." YouTube, 29 September 2026. https://www.youtube.com/watch?v=ArRaPV3QIqY
11. Kwa, Thomas, et al. "Measuring AI Ability to Complete Long Software Tasks." arXiv:2503.14499, first posted 18 March 2025. https://arxiv.org/abs/2503.14499
12. Bronars, Antonia, Younghyo Park and Pulkit Agrawal. "Tune to Learn: How Controller Gains Shape Robot Policy Learning." arXiv:2604.02523, April 2026. https://arxiv.org/abs/2604.02523
13. "Dyna Robotics closes $120M funding round to scale robotics foundation model." The Robot Report, 16 September 2025. https://www.therobotreport.com/dyna-robotics-closes-120m-funding-round-to-scale-robotics-foundation-model/