Full specifications
6 of 14 fields stated| Field | Value | Source and caveats |
|---|---|---|
| Release | ||
| Released | 13 July 2026 | [A] "Submitted on 13 Jul 2026" C |
| Input and output | ||
| Model type | world action model | Astribot's own term: "our next-generation latent world-action model". B |
| Inputs | camera language robot state | Technical report: "Conditioned on natural language instructions, visual observations, and robot proprioceptive states". The cameras are a head view and wrist views: "with the head view used to query wrist views". A |
| Controls | arms and hands whole body | Technical report: "guide a whole-body bimanual robot toward the desired outcome". Action groups: "actions are first decomposed into semantic groups corresponding to the torso and bimanual arms". A |
| Degrees of freedom controlled | not stated | |
| Size and speed | ||
| Parameters | not stated | Note The report gives no total count; it names its starting model: "by continuously training Qwen3.5-4B". A |
| Control rate | not stated | Note The report gives an inference time, not a control rate: "overall end-to-end inference time of 93.53 ms". A |
| Where it runs and what it is built on | ||
| On robot or cloud | not stated | Note Latency was measured "on a single NVIDIA RTX 5090 GPU"; the report does not say where the model runs on the robot. A |
| Base model | Qwen3.5-4B | Technical report, training recipe: "by continuously training Qwen3.5-4B". A |
| Openness | ||
| Access | not stated | |
| Licence terms | not stated | |
| Licence | not stated | |
| The developer’s own claims | ||
| Learns new tasks | needs retraining | Developer claim: "The model natively supports efficient fine-tuning with diverse, accessible data sources including egocentric human videos." The report's robot tasks use fine-tuning: "We adopt task-adaptive fine-tuning paradigms". B |
| Works across robots | not stated | Note Astribot's page says: "No specialized transfer learning algorithms are required to achieve cross-embodiment and human-to-robot skill transfer". The report's own robot tests use one robot: "All experiments are conducted on the high-performance bimanual mobile manipulation platform Astribot S1". 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
- Astribot presents Lumo-2 as the upgrade of its earlier model Lumo-1.
- Astribot published the Lumo-2 technical report on arXiv on 13 July 2026.
Versions and notes in the key figures (5)
Model type
Astribot's own term: "our next-generation latent world-action model".
Controls
Technical report: "guide a whole-body bimanual robot toward the desired outcome". Action groups: "actions are first decomposed into semantic groups corresponding to the torso and bimanual arms".
Inputs
Technical report: "Conditioned on natural language instructions, visual observations, and robot proprioceptive states". The cameras are a head view and wrist views: "with the head view used to query wrist views".
Parameters
Note The report gives no total count; it names its starting model: "by continuously training Qwen3.5-4B".
Base model
Technical report, training recipe: "by continuously training Qwen3.5-4B".
Sources
-
A
arxiv.
org/ · Retrievedhtml/ 2607. 11270 -
B
Primary source. www.
astribot. · Retrievedcom/ en/ AI/ Lumo2 -
C
arxiv.
org/ · Retrievedabs/ 2607. 11270
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.