AnswersPublished18 sources
What did Anyverse Dynamics announce in its October 22, 2026 release of the open-source MWA-Phoenix robot manipulation foundation model, including its 7-billion-parameter latent world action architecture, on-device operation using only RGB input without ground-truth depth, GitHub
Anyverse Dynamics announced MWA-Phoenix as an open-source, 7-billion-parameter latent world-action foundation model for robot manipulation, designed for on-device operation; coverage dates the announcement to October 22, 2026. [6][9] The available evidence confirms the broad release and some reporte Anyverse Dynamic...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Anyverse Dynamics announced MWA-Phoenix as an open-source, 7-billion-parameter latent world-action foundation model for robot manipulation, designed for on-device operation; coverage dates the announcement to October 22, 2026. [6][9] The available evidence confirms the broad release and some reporte
- Anyverse Dynamics announced MWA-Phoenix as an open-source, 7-billion-parameter latent world-action foundation model for robot manipulation, designed for on-device operation; coverage dates the announcement to October 22, 2026. [6][9] The available evidence confirms the broad rele
- ## Release and verified results
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Anyverse Dynamics announce in its October 22, 2026 release of the open source MWA Phoenix robot manipulation foundation model, incl. Article summary: Anyverse Dynamics announced MWA Phoenix as an open source, 7 billion parameter latent world action foundation model for robot manipulation, designed for on device operation; coverage dates the announcement to October 22,. Topic tags: general web, ai, automation, code, regulation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, chart
Anyverse Dynamics announced MWA-Phoenix as an open-source, 7-billion-parameter latent world-action foundation model for robot manipulation, designed for on-device operation; coverage dates the announcement to October 22, 2026. 6
9 The available evidence confirms the broad release and some reported results, but is insufficient to verify every technical detail and benchmark figure in your question.
Release and verified results
- Model and input: Reports describe a complete pretrained 7B model using RGB input without depth sensors; the announcement excerpt also says it does not use ground-truth depth.
11
14
- Code: The
Anyverse-Dynamics/Phoenix-MWAGitHub repository is available and identifies the work as “Learning Generalizable World–Action Models with Structural and Semantic Representations.” - Evaluation coverage: Announcement coverage names LIBERO-Plus, LIBERO-Pro, RoboCasa GR1, RoboTwin 2.0, RoboDojo real-robot testing, and Bench2Dex as the six evaluations on which the company claims leading performance.
10
- Scores supported by the excerpts: LIBERO-Plus 94.63%, LIBERO-Pro 66.48%, and RoboCasa GR1 Tabletop 82.67%.
Architecture and deployment components
- DINOv3: The company’s encoder notice confirms that its DINO110k checkpoint—including the backbone and adapter—is derived from Meta’s DINOv3.
- Phoenix-ILA and Phoenix-PhysCanvas: Retrieved coverage describes Phoenix-ILA as incorporating motion and contact into action learning, and Phoenix-PhysCanvas as an interactive simulated world model for cloud-based improvement before deployment.
- Unverified architecture details: Insufficient evidence to confirm VGGT geometry distillation, V-JEPA temporal-change distillation, the causal semantic VAE’s exclusion of future frames, jointly attending video/action experts, or asynchronous action-chunk execution.
Remaining evidence gaps and caveats
- I could not verify RoboTwin 2.0’s 72.42%, RoboDojo’s three bodies and 18 tasks, the 7.5-percentage-point gain over WALA, or the 17.3-point gain over pi0.5.
- Insufficient evidence to establish the specific remaining limitations of cross-body skill transfer, or to confirm pretraining from real-robot operation, egocentric human video and handheld UMI capture.
- I could not verify the claimed more than 200,000 hours of multimodal data, ongoing larger-model training, or the statement that no test data was replicated in training.
- The retrieved material presents benchmark leadership as announcement claims, not independent replication.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What did Anyverse Dynamics announce in its October 22, 2026 release of the open-source MWA-Phoenix robot manipulation foundation model, including its 7-billion-parameter latent world action architecture, on-device operation using only RGB input without ground-truth depth, GitHub"?
Anyverse Dynamics announced MWA-Phoenix as an open-source, 7-billion-parameter latent world-action foundation model for robot manipulation, designed for on-device operation; coverage dates the announcement to October 22, 2026. [6][9] The available evidence confirms the broad release and some reporte
What are the key points to validate first?
Anyverse Dynamics announced MWA-Phoenix as an open-source, 7-billion-parameter latent world-action foundation model for robot manipulation, designed for on-device operation; coverage dates the announcement to October 22, 2026. [6][9] The available evidence confirms the broad release and some reporte Anyverse Dynamics announced MWA-Phoenix as an open-source, 7-billion-parameter latent world-action foundation model for robot manipulation, designed for on-device operation; coverage dates the announcement to October 22, 2026. [6][9] The available evidence confirms the broad rele
What should I do next in practice?
## Release and verified results
Sources