Embodied AI Daily · 2026-10-13

Today's Highlights

• Bloomberg: Liang Wenfeng is holding intensive talks with embodied AI founders; DeepSeek is considering entering the field directly or via investment
• DeepSeek reportedly plans to lead a tens-of-millions-of-dollars seed round in Winverse AI, founded by Swin Transformer co-author Hu Han, at a ~$400M post-money valuation
• Shanghai launches a new-generation smart manufacturing action plan: 8,000 humanoid robots into factories by 2028
• Huxiu survey: the mass-production year is not the commercialization year — industry capital still circulates within the system
• Three automaker paths in one day: BYD puts robots in showrooms not workshops; Toyota revives full-scope robot R&D; XPeng targets year-end mass production and H1 2027 deliveries
• Concrete numbers for robot memory demand: DRAM from ~8,300 Tbit in 2026 to ~172,000 Tbit by 2030

Details

1. DeepSeek weighs entering embodied AI; first surfaced investment is about buying people

• Source: Bloomberg (with follow-ups from Huxiu, Wallstreetcn, ifeng, NetEase, etc.)
• Timing: circulating since the afternoon of 2026-10-12, continuing on 10-13
• Bloomberg reports that Liang Wenfeng has recently been in intensive contact with embodied AI founders, evaluating whether DeepSeek should build a team directly or enter through investment. The first surfaced target is Winverse AI: founder Hu Han is one of the authors of Swin Transformer and previously served as chief scientist of Tencent's vision-language and embodied foundation model team before founding the company in 2026. DeepSeek reportedly plans to lead a seed round in the tens of millions of dollars at a post-money valuation of about $400 million, with Meituan's Long-Z fund committed to participate; sources say the deal is not finalized and details may change. Additional context: per NetEase, DeepSeek previously subscribed roughly 2% of Unitree Robotics' shares at its Shanghai IPO — from secondary-market equity to seed-round lead, Liang has now made two moves in physical intelligence.

2. Shanghai: 8,000 humanoid robots into factories by 2028

• Source: Cailian Press / 21 Finance / Shanghai Observer
• Shanghai launched an action plan to accelerate new-generation smart manufacturing, explicitly targeting 8,000 humanoid robots deployed in factories by 2028. Shanghai Observer's same-day headline framed the industry's current position: embodied intelligence moving from "can move" to "can work." This is the first local-government target with both a unit count and a deadline, following top-level policy documents listing humanoid robots as a "new-generation intelligent terminal."

3. Huxiu commercialization survey: the mass-production year is not the commercialization year

• Source: Huxiu / Yicai (First Financial)
• Huxiu's survey of embodied AI commercialization concludes: this year's mass production does not equal commercialization; industry capital still circulates within the system — orders sold to each other, rounds invested in each other — while revenue from external customers is not yet the mainstay. A counterpoint also appeared: a Yicai discussion piece argues "GPT-7 may break through the embodied intelligence problem," potentially accelerating the robot boom; another Huxiu piece placed Wang Xingxing's "embodied intelligence tipping point is moving earlier" claim alongside Liang Wenfeng's moves — the tipping-point timeline is being pulled forward by the industry's best accountants.

4. Three automaker approaches: showrooms, full lineups, and rushing to mass production

• Source: Robot Lecture Hall / Nikkei China / Sina
• BYD's direction is clearer: its humanoid robots go into showrooms, not workshops — targeting dealer and sales scenarios, reusing automaker channel and scene assets. Toyota, per Nikkei China, is again challenging the world with "full-scope robot R&D," against the backdrop of a reported factory rebuild plan: roughly ¥1 trillion (about $6.4B) per year from 2028 to rebuild plants and deploy large-scale robots working alongside people. XPeng reaffirmed its timeline: humanoid robot mass production by the end of this year, deliveries starting in the first half of 2027. Three automakers, three tempos — showrooms, factories, deliveries.

5. Robot memory demand gets concrete numbers: 20x in four years

• Source: Guandian / foreign media (institutional estimates)
• Institutional estimates: humanoid robots' DRAM demand will grow from ~8,300 Tbit in 2026 to ~172,000 Tbit in 2030 — more than 20x in four years — with Samsung and SK hynix seen as the main beneficiaries. The 8,300 Tbit base means robot memory is already a real shipping market, not a distant assumption; the 20x slope puts a price on the claim that "every robot needs large local memory."

6. Junpu Intelligent launches full-stack industrial embodied AI platform Primus Stack

• Source: Eastmoney / Sohu Finance
• Junpu Intelligent officially released Primus Stack, a full-stack industrial embodied intelligence platform covering the entire chain from robot body and models to production-line integration. The same day, Shanghai hosted the "Industrial Embodied Intelligence, Unbounded Manufacturing" industry innovation summit. Industrial scenarios are becoming the most concentrated track for embodied AI deployment — ahead of home and service scenarios, with customer willingness to pay maturing earlier.

7. Geek+ CTO: the next phase is about stability under extreme conditions

• Source: Sina Finance
• Geek+ CTO Li Hongbo on embodied intelligence's next phase: the competition is not more demos but reliability under extreme conditions — no failures in high-temperature, high-load, continuous warehouse operations. The shift from "capability showcase" to "engineering reliability" echoes Huxiu's commercialization survey: whether a robot can work is ultimately decided by the longest shift it can stand.

Paper Corner (arXiv cs.RO new batch)

• MC-EvoBench (arxiv.org/abs/2610.13829): an exam for "experience-driven self-improvement." Registers 19 evolution chains in a controlled Minecraft environment; core claim: having memory and completing tasks does not equal acquiring reusable capabilities; evaluation must separate capability change from carried-over resources, data-exposure asymmetry, and early execution failures.
• LAPT-Scale (arxiv.org/abs/2610.13732): robot interaction data is expensive; human egocentric video is a free mine. Learns continuous latent actions to explain inter-frame visual changes, turning human video into supervision signals, then transferring to real-robot policies — another answer to the data-cost problem.
• SmolWAM (arxiv.org/abs/2610.13740): a world-action model on edge devices. Keeps the pretrained 2B-parameter Cosmos Policy backbone, retains only 12 of 28 layers, and uses single-frame visual latents with cached text features to bring "predicting the future" into a robot body's compute budget. Adding to yesterday's LeWAM/MiniWAM, the world-action model series gains another entry.
• JitterVLA (arxiv.org/abs/2610.13628): a replay benchmark for VLA action quality. Feeding the same instructions and arrival timing to different controllers isolates the controller's effect; interpolation reduces modeled p99 jerk from 615 to 106 m/s³ (an 83% reduction), using 20 human hand-trajectory recordings, 838 data segments across three open-source VLA pipelines, and one physical UR3 — the first reproducible yardstick for action smoothness in robots next to humans.

Editor's Note

The biggest signal today is money changing hands. Bloomberg reports Liang Wenfeng is in intensive contact with embodied AI founders, with DeepSeek weighing direct entry or investment; the first surfaced target is Winverse AI, founded by Hu Han, at tens of millions of dollars seed round and roughly $400M post-money. What makes this deal interesting is what is being bought: not robots, but people. Hu Han co-authored Swin Transformer and previously led vision-language and embodied foundation model work at Tencent. For a large-model company built on training efficiency, giving its first major embodied investment to a vision-Transformer author suggests the price is not priced on revenue but on "if the next generation of large models really cracks embodied intelligence, whose brain gets loaded into robots first." The deal is not finalized and details may change — Bloomberg's sourcing convention is anonymous informed sources, so read with a grain of salt.

On the other side is demand-side data. Shanghai set a deadline-bound target: 8,000 humanoid robots into factories by 2028. BYD's robots go into showrooms, not workshops. Toyota is reviving full-scope robot R&D. Geek+ says the next phase is stability under extreme conditions. Put the $400M valuation next to the 8,000-unit target: one is priced on "possibility," the other on "delivery" — the "two prices coexisting" noted yesterday has now extended from the robot market to the model side. Huxiu's survey adds a third sentence: the mass-production year is not the commercialization year, and industry capital still circulates within the system. DeepSeek's investment is arguably system money too, but at least it buys people, not stories. MC-EvoBench in the paper corner is a fitting footnote: while capital places the tipping point next year, academia released an exam the same day testing "how to prove capabilities actually improved" — stored memory and completed tasks don't count as learning.