# The week physical AI went open — and left the US labs out of the frame

URL: https://www.thedeepfeed.ai/posts/2026-07-11-the-week-physical-ai-went-open/
Category: Research
Published: 2026-07-11
Author: the-deep-feed
Tags: robotics, world-models, embodied-ai, open-weights, china, physical-ai
Kind: deep

> In five days, an Ant Group affiliate open-sourced an entire robot brain — perception, depth, world model, video-action — while BAAI shipped a world model that learns robot control without action labels and Mistral entered robotics for the first time. The physical-AI frontier moved to open weights, and almost none of it came from a US lab. A map of the week that reset embodied AI.

## TL;DR

- In five days (**Jul 7–11**), the physical-AI frontier moved to **open weights** — and almost none of it came from a US lab. **Robbyant** (Ant Group's embodied-AI unit, the company behind **Alipay**) shipped a **six-model full stack** for robots; **BAAI** released **Orca**; **Mistral** entered robotics for the first time with **Robostral Navigate**.
- The headline is Robbyant's **LingBot-VA 2.0**, billed as *"the industry's first embodied-native video-action world model"* — pretrained from scratch for physical control instead of fine-tuning a digital video generator. It leads RoboTwin 2.0 (**93.6 avg** vs π0.5's 79.8) on the four demo tasks: air hockey, chip picking, conveyor sorting, desk tidying.
- **BAAI Orca** is the research shock: a *"world foundation model"* that **matches π0.5 across five manipulation tasks even though its base model never saw a single action label** in pretraining. If it holds up, it's a partial answer to robotics' chronic data-scarcity problem.
- The licensing is a mixed map, and it matters. **LingBot-Vision** is genuinely open (**Apache-2.0**); **LingBot-World 2.0** is research-only (**CC-BY-NC-SA**); **Mistral's Robostral is NOT open-weight** at all (API/enterprise). "Physical AI went open" is the trend, not a blanket fact — read the license before you build.
- The strategic read: this connects to [the gate and the giveaway](/posts/2026-07-04-the-gate-and-the-giveaway/). The US frontier gate is aimed at *language-model* cyber-risk. The *embodied* frontier — the one that touches factories, logistics, and humanoids — is being seeded in open weights, mostly from China, while US labs sit outside this week's frame entirely.

The AI story everyone watched in early July was a language model that wouldn't ship: [GPT-5.6 Sol](/posts/2026-07-09-the-voluntary-gate-that-works-like-a-license/), held twelve days behind a US government review. But the more consequential releases of the same week weren't language models at all, and no government reviewed any of them. Between July 7 and July 11, a cluster of **world models, vision-action models, and embodied-AI foundation models** landed — mostly open weights, mostly from China, and almost entirely outside the US labs that dominate the chat-model conversation.

One of them, an Ant Group affiliate, open-sourced what amounts to an entire robot brain in five days. This is a map of that week, and of the frontier it just redrew.

# The releases, in order

| Date | Release | Org (country) | What it is | Open? |
|---|---|---|---|---|
| Jul 7 | **LingBot-Vision** | Robbyant / Ant Group (CN) | 1B boundary-centric self-supervised ViT — dense spatial perception, "the eyes" | 🟢 Apache-2.0 |
| Jul 7–8 | **LingBot-VLA 2.0** | Robbyant / Ant Group (CN) | Open-sourced vision-language-action "universal brain" | 🟢 Open-sourced |
| Jul 8 | **LingBot-World 2.0 (Infinity)** | Robbyant / Ant Group (CN) | Interactive world model — hour-long real-time generation at 720p/60fps | 🟡 CC-BY-NC-SA (non-commercial) |
| Jul 9 | **Robostral Navigate** | Mistral AI (FR) | 8B single-RGB-camera navigation, 76.6% on unseen R2R-CE; Mistral's first robotics model | 🔴 API/enterprise, no public weights |
| Jul 10 | **LingBot-VA 2.0** | Robbyant / Ant Group (CN) | "Industry's first embodied-native video-action world model," causal DiT pretrained from scratch | 🟡 Paper + project page |
| Jul 11 | **Orca** | BAAI (CN) | World foundation model matching π0.5 across 5 tasks with no action labels in pretraining | 🟢 Technical report / open research |

*(Robbyant also shipped LingBot-Depth 2.0 and LingBot-Video in-week, completing a six-model stack — [per Startup Fortune](https://startupfortune.com/robbyant-builds-out-an-embodied-native-full-stack-for-physical-ai/).)*

Read the "Org" column. Four of the six named releases come from a single Chinese company. One more comes from a Chinese government-backed nonprofit. The lone non-Chinese entrant, Mistral, is French — and is the one release with no open weights and a press narrative of playing catch-up. The US frontier labs that gate the language-model conversation are not in this table at all.

# First, the vocabulary — because the labels are the argument

This cluster is easy to wave off as "more robot models." It isn't, and the distinctions are the whole story. Three terms are doing the work.

A **vision-language-action (VLA)** model is reactive: it takes an image and a text instruction and outputs robot actions, usually by fine-tuning a vision-language model on robot trajectories. See the surroundings, read the command, move. It does not model what happens *next*.

A **world model** adds prediction: it learns how the environment evolves, so it can imagine the consequence of an action before taking it. The academic framing calls the merged category **World Action Models** — [in the words of a June 2026 survey](https://arxiv.org/abs/2605.12090), *"embodied foundation models that unify predictive state modeling with action generation, targeting a joint distribution over future states and actions rather than actions alone."*

The whole point of this week is that the field moved decisively from the first toward the second. The most-hyped releases weren't better instruction-followers. They were models that **build an internal picture of the physical world and act inside it** — and two of them claim to do it without the thing everyone assumed was the bottleneck: labeled robot data.

# Robbyant: an entire robot brain in five days

The center of the week is Robbyant, the embodied-AI unit inside Ant Group — the fintech that runs **Alipay**. That a payments company is building robot foundation models is itself the tell of how contested this frontier has become. Its own mission statement is unambiguous about scope: *"building one brain for all robots … a full-stack technology spanning spatial perception, action models, and environmental reward."*

Over five days it made good on "full-stack" literally, shipping the pieces of a robot brain one at a time. [Startup Fortune](https://startupfortune.com/robbyant-builds-out-an-embodied-native-full-stack-for-physical-ai/) captured the ecosystem-map thesis better than any single release could:

> In the span of a single week, Robbyant has released a cluster of models that, taken together, amount to something larger than any one of them on its own: a complete, ground-up stack for embodied artificial intelligence … each built specifically for robots operating in the physical world rather than adapted from tools designed for digital content.

The two anchors of that stack are worth reading closely.

**LingBot-World 2.0 (Infinity)** is an interactive world model that generates a playable environment and — the headline claim — keeps it coherent for an *hour* of continuous real-time generation at 720p and 60fps with sub-second latency. Where most video-based world models "blur and collapse" over long horizons, Robbyant claims an *"unbounded interaction horizon while maintaining consistent output quality."* [The Next Web put it vividly](https://thenextweb.com/news/robbyant-ant-group-lingbot-embodied-ai-open-source): *"The company behind China's Alipay just gave away an AI that builds a playable video game world and keeps it running for a full hour … For a company most people know through a payments app, it is a bold land-grab in embodied AI."*

**LingBot-VA 2.0**, shipped July 10, is the more technically aggressive claim: *"the industry's first embodied-native video-action world model."* The pitch is a direct shot at how the rest of the field builds these systems.

> Most mainstream approaches rely on video generation models designed for digital content, which are then fine-tuned for robot control. Because content creation prioritizes visual quality and creativity, while robot control requires execution efficiency and physical accuracy, this forced adaptation often leads to knowledge forgetting and reduced generalization.
>
> — [Business Wire China](https://www.businesswirechina.com/en/news/63483.html), July 10, 2026

Robbyant's alternative is to pretrain a causal video-action model *from scratch*, natively for embodiment — so the model, in the release's words, is *"designed to understand how an action will change the environment and to decide the next step based on that causal prediction."* On the RoboTwin 2.0 benchmark, [MarkTechPost reports](https://www.marktechpost.com/2026/07/11/ant-groups-robbyant-unveils-lingbot-va-2-0/) LingBot-VA 2.0 averaging **93.6** against π0.5's 79.8 and X-VLA's 72.9, across a demo suite [Robotics 24/7 lists](https://www.robotics247.com/article/robbyant-launches-lingbot-va-2.0-embodied-native-world-action-model) as *"fast-paced air hockey and fragile chip picking to conveyor-belt sorting and long-horizon desk tidying."*

The "native vs fine-tuned" argument is the load-bearing idea of the whole week. It says the shortcut the field took — repurpose a video generator, bolt on actions — was a dead end, and that physical AI needs models grown for physics from the first token. Whether that's true is the bet. That an Alipay affiliate is the one making it, in open weights, is the surprise.

# BAAI's Orca: the data-scarcity heresy

If Robbyant supplied the week's ambition, BAAI supplied its genuine research shock. On July 11, the Beijing Academy of Artificial Intelligence — a government-backed nonprofit best known for the WuDao models and BGE embeddings — released **Orca**, and the claim is the kind that resets a subfield if it survives replication.

> BAAI's Orca world model matches the performance of specialized systems across five robotics tasks, even though its base model was trained without a single action label. The approach could help solve robotics' chronic data shortage.
>
> — [The Decoder](https://the-decoder.com/chinas-orca-world-model-matches-specialized-robotics-systems-without-ever-seeing-a-single-action-label/), July 11, 2026

Sit with why that matters. The central constraint in robotics is that action-labeled data — recordings of a robot doing a task, annotated with the exact motor commands — is brutally expensive to collect. It is the field's equivalent of the data wall that language models hit, except worse, because you can't scrape it off the web. Orca's design attacks that wall directly. Its base model learns from raw, unlabeled video in a mode BAAI calls *"unconscious learning"* — *"the model sees an image and predicts what the next one will look like, not at the pixel level, but in an abstract space, picking up motion patterns, occlusions, and typical scene dynamics."* Only a small "Action Expert" head, trained on as few as 200 recordings per task, connects that internal world-picture to motor output.

The result: on five manipulation tasks with a two-armed humanoid — shelving books, stacking bowls, scooping sugar — Orca **matches π0.5, a system built specifically on robot data**, despite its base model never having seen "which movement goes with which image." If that generalizes, the economics of training robots change: the expensive labeled data becomes a thin finishing layer on top of cheap, abundant video. That is the single most important research claim in the cluster, and it came from a nonprofit, released as open research.

# Mistral: the West's lone entry, and the caveat that proves the pattern

The one non-Chinese release is also the one that most clearly marks the asymmetry. Mistral's **Robostral Navigate**, its first embodied model, is a genuinely elegant piece of engineering: an 8B model that navigates from a *single ordinary RGB camera*, no LiDAR, no depth sensors, and still hits *"76.6% on R2R-CE … beats the best single-camera approach by 9.7 points and the best system using depth or multiple cameras by 4.5 points, despite using neither."* Trained entirely in simulation, it generalizes across wheeled, legged, and flying robots.

But two things about it underline the week's real story rather than softening it. First, the framing: [InfoWorld's headline](https://www.infoworld.com/article/4195667/mistral-joins-rush-to-develop-ai-for-robots.html) was "Mistral joins rush to develop AI for robots," noting that "other AI model developers are ahead of the game." The West's most prominent open-weight lab is entering embodied AI as a *latecomer*. Second, and more telling: Robostral is the one release in the cluster with **no open weights**. Mistral — the company whose entire brand is open models — shipped its robotics debut as an API/enterprise product. The lab that made its name giving language models away is holding its robot model close, in the same week a Chinese payments affiliate gave a whole robot stack away.

# Read the license before you read the hype

"Physical AI went open" is the trend of the week, but it is not a blanket fact, and treating it as one will burn anyone who builds on it. The licensing is a genuine patchwork:

- 🟢 **LingBot-Vision** is the real thing — Apache-2.0 on Hugging Face, four sizes, commercial use permitted. The most permissively licensed model in the cluster.
- 🟡 **LingBot-World 2.0** ships its 14B variant under **CC-BY-NC-SA-4.0** — open for research, *non-commercial*, share-alike. "Open-sourced" in the press release; not free for a product.
- 🟡 **LingBot-VA 2.0** and **BAAI Orca** released papers and project pages; downloadable weights and licenses should be verified in each repo before you assume you can ship on them.
- 🔴 **Robostral Navigate** is not open-weight at all.

The distinction between "open research" and "open weights you can build a company on" is exactly the kind of detail that gets flattened in a busy news week. On this frontier, flattening it is expensive.

# Why this matters

The tidy read is that this was a big week for robots. The sharper read is that it was a big week for *where the embodied frontier is being built* — and the answer is: increasingly in open weights, increasingly from China, and this week not from the US labs at all.

That connects directly to the policy thread we've been pulling. The [US frontier gate](/posts/2026-07-09-the-voluntary-gate-that-works-like-a-license/) is aimed at language-model cyber-risk — the capability class of 2025. The [open-weight counter-move](/posts/2026-07-04-the-gate-and-the-giveaway/) we mapped over the July 4 weekend was about coding models shipped free on domestic chips. This week adds the third front, and arguably the one with the longest tail: the *embodied* frontier — the models that will eventually run factories, warehouses, logistics fleets, and humanoids — is being seeded in the open, largely by Chinese labs, while the American frontier conversation stays fixated on chatbots and their clearance paperwork.

There's a through-line from [Jim Fan's thesis](/posts/2026-05-01-jim-fan-robotics-llm-playbook/) that robotics is retracing the LLM playbook. If that's right, this week was the field's DeepSeek moment for physical AI: the point where the open-weight, cost-efficient, non-US challengers stopped trailing the paradigm and started defining it. The world models that imagine consequences, the action models that learn without labels, the full stacks given away by companies you'd never have guessed were building robots — that is what the frontier looks like when no one is asking permission to ship it.

The language-model frontier spent early July learning it now needs a government's clearance. The physical-AI frontier spent the same week proving it doesn't intend to ask.

## Sources

- [Startup Fortune — Robbyant Builds Out an Embodied-Native Full-Stack for Physical AI (Jul 10, 2026)](https://startupfortune.com/robbyant-builds-out-an-embodied-native-full-stack-for-physical-ai/)
- [The Next Web — Ant Group's Robbyant open-sources embodied-AI stack (Jul 9, 2026)](https://thenextweb.com/news/robbyant-ant-group-lingbot-embodied-ai-open-source)
- [Business Wire China — Robbyant Unveils LingBot-VA 2.0 (Jul 10, 2026)](https://www.businesswirechina.com/en/news/63483.html)
- [MarkTechPost — Ant Group's Robbyant Unveils LingBot-VA 2.0 (Jul 11, 2026)](https://www.marktechpost.com/2026/07/11/ant-groups-robbyant-unveils-lingbot-va-2-0/)
- [MarkTechPost — Ant Group's Robbyant Open-Sources LingBot-Vision (Jul 7, 2026)](https://www.marktechpost.com/2026/07/07/ant-groups-robbyant-open-sources-lingbot-vision/)
- [MarkTechPost — Meet LingBot-World Infinity, an open causal world model with an agentic harness (Jul 9, 2026)](https://www.marktechpost.com/2026/07/09/meet-lingbot-world-infinity-an-open-causal-world-model-with-an-agentic-harness/)
- [The Decoder — China's Orca world model matches specialized robotics systems without ever seeing a single action label (Jul 11, 2026)](https://the-decoder.com/chinas-orca-world-model-matches-specialized-robotics-systems-without-ever-seeing-a-single-action-label/)
- [Mistral AI — Robostral Navigate (Jul 9, 2026)](https://mistral.ai/news/robostral-navigate/)
- [InfoWorld — Mistral joins rush to develop AI for robots (Jul 10, 2026)](https://www.infoworld.com/article/4195667/mistral-joins-rush-to-develop-ai-for-robots.html)
- [Robotics 24/7 — Robbyant launches LingBot-VA 2.0 embodied-native world-action model (Jul 11, 2026)](https://www.robotics247.com/article/robbyant-launches-lingbot-va-2.0-embodied-native-world-action-model)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-07-11-the-week-physical-ai-went-open/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt