Full specifications
5 of 14 fields stated| Field | Value | Source and caveats |
|---|---|---|
| Release | ||
| Released | 29 September 2026 | [A] "September 29, 2026" A |
| Input and output | ||
| Model type | not stated | Note Dyna's own term: "Dyna-2.1, a full-stack physical agent that can autonomously complete end-to-end real-world workflows". Its action layer: "The Dyna-2 policy, an improved version of our Dyna-2 world-action model". A |
| Inputs | camera language | Stated for the workflow layer only: "Camera + text memory"; "People can also give it short instructions, which it expands into steps the policy can carry out". The policy's own inputs are not stated. A |
| Controls | arms and hands whole body | The Dyna-2 policy "turns the current step into whole-body target trajectories to be tracked by the whole-body controller". A |
| Degrees of freedom controlled | not stated | Note The blog labels the robot, not the model's controlled joints: "21-DoF upper body + mobile base". A |
| Size and speed | ||
| Parameters | not stated | |
| Control rate | 100Hz | The fastest stated layer: the whole-body controller turns trajectories "into joint targets and wheel velocities at 100 Hz". The policy layer is labelled "world-action model · ~5 Hz". A |
| Where it runs and what it is built on | ||
| On robot or cloud | not stated | |
| Base model | not stated | Note The policy layer: "The Dyna-2 policy, an improved version of our Dyna-2 world-action model". The workflow layer is "built on a vision-language model pre-trained on diverse text and images". A |
| Openness | ||
| Access | not stated | |
| Licence terms | not stated | |
| Licence | not stated | |
| The developer’s own claims | ||
| Learns new tasks | needs retraining | Developer claim: Dyna-2.1 "learns new skills much faster than our previous systems, with far less demonstration data". Training: "We pre-train the action model, an improved Dyna-2 policy, on one million hours of human video mixed with robot data from our fleet"; "so fine-tuning on robot data only has to teach the policy Taku’s specific body". A |
| Works across robots | not stated | Note Dyna describes its data format as embodiment-agnostic, not the model as running on other robots: "The same format describes a human, humanoid, semi-humanoid or tabletop bimanual robot". A Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count. |
How to read the figures
- 50HzA
- The letter in the grey square after a figure points to the source the figure came from. Pages with several sources carry several letters.
Notes
- Dyna Robotics introduced Dyna-2.1 in a blog post dated 29 September 2026.
- Dyna-2.1 is a stack of three model layers: a whole-body controller, the Dyna-2 policy and a workflow layer.
Versions and notes in the key figures (4)
Model type
Note Dyna's own term: "Dyna-2.1, a full-stack physical agent that can autonomously complete end-to-end real-world workflows". Its action layer: "The Dyna-2 policy, an improved version of our Dyna-2 world-action model".
Controls
The Dyna-2 policy "turns the current step into whole-body target trajectories to be tracked by the whole-body controller".
Inputs
Stated for the workflow layer only: "Camera + text memory"; "People can also give it short instructions, which it expands into steps the policy can carry out". The policy's own inputs are not stated.
Base model
Note The policy layer: "The Dyna-2 policy, an improved version of our Dyna-2 world-action model". The workflow layer is "built on a vision-language model pre-trained on diverse text and images".
Sources
-
A
Primary source. www.
dyna. · Retrievedco/ dyna-2. 1
One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.