Ultra raises $62M on Physical Intelligence's brains, RobCo turns unicorn, and open coding models get bigger
Ultra raised $62M (Oct 9) for warehouse robots it leases by the month, and it runs them on Physical Intelligence's AI models instead of building its own. RobCo, which has offices in Munich, San Francisco and Austin, passed a $1B valuation (Oct 5), double its January mark. UBTECH and FAW-Volkswagen signed a deal on Oct 8 to test humanoids in car-plant logistics. Open-weight coding models got bigger: Reflection announced the 501B-parameter Beam, and IQuest released the 320B IQuest-Q1, both aimed at coding agents.
The week at a glance
- Ultra: $62M in total (a $50M Series A led by Framework Ventures plus a $12M seed), monthly robot leases, and AI models from Physical Intelligence. 1
- RobCo: valued above $1B after new investment and an employee share sale, double its valuation nine months earlier. 9
- UBTECH: the FAW-Volkswagen partnership moves from vehicle inspection to factory logistics. 3
- Open-weight coding models: Reflection's Beam (501B total, 23B active) and IQuest-Q1 (320B total, 15B active). 1415
- Austin hiring: Amazon Robotics lists Austin on 89 open roles, Apptronik on 65 and Tesla Optimus on 43. All three counts are unchanged from last week.
Companies in this issue
1. Robots at work: warehouses, car plants and supermarkets
Ultra raises $62M and pairs its robots with Physical Intelligence models
- The money: a $50M Series A led by Framework Ventures, with Y Combinator participating, on top of an earlier $12M seed led by Y Combinator and NextView. Ultra is based in Brooklyn and announced both rounds on Oct 9. 12
- The product: a two-armed torso mounted on an industrial arm that packs e-commerce orders at third-party logistics warehouses. Ultra says its robots have packed more than 500,000 orders.
- The business model: customers pay an up-front integration fee and then a monthly fee for hardware and software support. CEO Jon Miller Schwartz told Fortune that demand has let Ultra raise its prices. He did not disclose revenue.
- The AI: Ultra builds and installs the robots, and Physical Intelligence (valued at $5.6B, per Fortune) supplies the models that adapt them to each customer's site. Physical Intelligence in turn gets real-world data from working warehouses. Schwartz expects humanoids to reach this kind of scale in about five years.
UBTECH and FAW-Volkswagen move humanoid tests into factory logistics
- On Oct 8, UBTECH and FAW-Volkswagen, the Chinese joint venture that builds Volkswagen, Audi and Jetta cars, signed a strategic cooperation agreement to develop and test humanoid robots for logistics inside car plants.
- It builds on earlier trials at FAW-Volkswagen's Qingdao plant, where UBTECH's Walker S Lite took part in vehicle quality inspection.
- Neither company has disclosed specific workflows, robot counts or performance results. 3
Also this week
- Humanoids restocking Danish supermarket shelves at night: HIVE Robots, a Danish startup, is training robots built by Unitree at a Føtex store near Copenhagen. Føtex is part of Salling Group, Denmark's largest retailer. Human operators still guide every movement. The target is independent overnight restocking by early 2027, and Føtex says staff spend 40–50% of their time refilling shelves. Danish law would have to change before the robots could work among shoppers. 45
- DEEP Robotics ran field trials of its DR02 humanoid in Hangzhou's Xihu District in early October. The robots tested traffic control at intersections and answered visitors' questions in several languages. 6
- Jabil, the contract manufacturer that works with Apptronik on its Apollo humanoid, says humanoids are entering a phase of "tens of thousands" of units, when hand assembly no longer works and cost, manufacturability and quality take over. It also named edge-compute prices as an ongoing obstacle to scale. 7
- Tesla Optimus (context): The Information reported on Sept 25 that Tesla built "several hundred" Optimus robots a week in August, mostly for internal training. Tesla has not confirmed any unit counts. Its Q3 results on Oct 21 are the next chance for official figures. 8
2. Funding: RobCo turns unicorn, and AI agents keep raising
- RobCo (industrial robots for factories) passed a $1B valuation on Oct 5, double its valuation nine months earlier. The deal combined new investment with a share sale by employees, and the size of each part was not disclosed. Sequoia, Lightspeed, Greenfield, Kindred, Lingotto and Promus joined new investors Cherry Ventures and European Tech Collective. RobCo raised a $100M Series C in January. The press release is datelined San Francisco, Austin and Munich. 910
- Manus, the Chinese AI-agent company, raised more than $500M, led by Boyu Capital and IDG Capital. It is the company's first round since Chinese authorities unwound Meta's planned acquisition. The valuation was not disclosed. 11
- Arena (formerly LMArena), which runs the crowd-voted AI model leaderboard, raised a $200M Series B at a $3.1B valuation, led by Lightspeed and Khosla. The company says it reached $100M in annualized revenue in June, and it has added an alignment leaderboard. 12
- PONGBOT, a Chinese maker of table-tennis and tennis training robots, closed a Series B worth "hundreds of millions of RMB," led by the Shanghai State Investment Pilot Fund. 13
3. Models & tools: open-weight coding models get bigger
- Reflection AI: Beam (Oct 5). Its first open-weight model is a mixture-of-experts with 501B total parameters and 23B active, trained on 23.8T tokens. Reflection reports 80.9% on SWE-bench Verified and 80.1% on Terminal-Bench 2.1, close to GLM 5.2 and behind Kimi K3. It says Beam uses 3–4× less inference compute than GLM 5.2 on reasoning tests. Access is by waitlist for now, and Apache 2.0 weights are promised later in October. 14
- IQuest-Q1 (Oct 9). IQuest Research released open weights on Hugging Face for a 320B mixture-of-experts model with about 15B active parameters and a 524K-token context. It is built for coding agents and command-line work. IQuest itself calls the model early-stage and says it needs human oversight. 15
- Microsoft-Decision-1 (Oct 9) is a fast model for routing, classification, prioritization and verification steps inside agents and workflows. It is available in Microsoft Foundry and on OpenRouter. 16
- Xiaomi MiMo-V2.6 technical report (Oct 8): a family of models that take text, images and other inputs, trained with heavily scaled reinforcement learning at contexts of up to 1M tokens. 17
- Anthropic (Oct 8) launched the Anthropic Cyber Mission. It gives critical-infrastructure defenders frontier models and on-site engineers through 11 founding partners, including CrowdStrike, Palo Alto Networks and Rockwell Automation, and adds a free open-source vulnerability scanner. Anthropic also pledged $150M over three years to the federal Genesis Mission for AI-driven science. 1819
4. Research & open source: robots that remember and navigate
- Long-WAM (NVIDIA with MIT, HKU and UCSD, arXiv, Oct 7) was among the most upvoted robotics papers on Hugging Face Daily Papers this week. It shows that giving a world-action model more visual history pays off when the video model is pretrained autoregressively. On RoboCasa GR-1, success rose from 63.3% to 78.7% as context grew from 0 to 19.2 seconds. The authors report it runs on an RTX 5090, DGX Spark and Jetson AGX Thor, and the code is open. 2021
- SuperNav (Zhejiang University, Oct 8) leaves a general multimodal model unchanged and gives it navigation skills and tools. The model picks destinations by pointing at locations in camera images, and no navigation-specific fine-tuning is needed. Code is open. 2223
- DreamTrue (Oct 8) is a robot world model that is trained on deliberately altered actions, including failures, so its video predictions follow the commanded action instead of drifting toward success. It reports state-of-the-art action following on AgiBot data. 24
- SimpleICL is a minimal, low-cost recipe for robots that learn a task from a visual demonstration shown at run time, with no large-scale pretraining. The authors plan to open-source the data and training code. 25
Note: citation counts can't be measured for papers this new. These picks are based on community attention (Hugging Face Daily Papers upvotes) and how relevant they are to humanoid development.
5. Hiring and the map: Austin holds steady
- A quiet week for hiring. Open roles at the six companies we track went from 1,001 to 1,006. No postings closed and no hiring trend started or faded. This comparison covers only 3–4 days, so next week's is the first full week.
- Amazon Robotics added 2 roles, both research jobs in its Frontier AI & Robotics group in San Francisco, which now has 17 openings. The group's hardware, firmware and lab-manager roles may point to an in-house robot platform (low confidence).
- Apptronik added a global supply manager for electrical and electronics parts and a configuration manager. That may indicate early supplier build-out ahead of production (low confidence, two postings).
- Figure +2 (security and accounting) and Agility +1 (business systems). Tesla Optimus (291) and Physical Intelligence (38) were flat.
- The map grew from 67 to 82 companies since Issue #1. New entries include Austin's TerraFirma and Contoro Robotics, Amazon Robotics, and industrial robot makers such as RobCo, Sereact, Wandelbots, Mujin and Telexistence. The supply-chain section now lists 45 component suppliers. 26
6. What to watch next week
- Open weights from Reflection (Beam) and Mistral (Large 4), both promised by the end of October.
- RoboBusiness 2026, Oct 20–21 in Santa Clara, and Tesla's Q3 results on Oct 21, the next chance for official Optimus figures.
- Whether more robot makers announce Physical Intelligence partnerships like Ultra's, and whether RobCo follows its share sale with a new primary round.
- Agility's SPAC close (expected ticker AGLT) and the still-unconfirmed reports of an Nvidia investment in Figure.
Method: items were checked against primary or dated sources and limited to announcements from Oct 5–12, 2026 that Issue #1 did not cover. Context items from outside that window are labeled. Charts use only disclosed figures from the linked sources. Company claims (revenue, benchmarks, order counts) are attributed to the company and have not been independently verified. Hiring counts come from public job boards.
Sources
- Fortune, “Ultra raises $62 million for fast-growing ‘robots as a service’ business, announces tie-up with AI research firm Physical Intelligence”, Oct 9 ↩
- FinSMEs, “Ultra Robotics Raises $50M in Series A Funding” ↩
- A3 Industry Insights, “UBTECH and FAW-Volkswagen: From Humanoid Robot Inspection to Factory Logistics”, Oct 12 ↩
- Euronews, “Denmark tests humanoid robots in supermarkets to help to ease staff workload at night”, Oct 11 ↩
- HIVE, “HIVE and føtex Enter Co-Innovation Partnership to Pilot Humanoid Robots in føtex Stores”, March partnership announcement ↩
- DEEP Robotics via ACCESS Newswire, “DEEP Robotics' Humanoid Robot Debuts in Xihu District, Exploring New Approaches to Smart Transportation”, Oct 8 ↩
- The Robot Report, “Jabil discusses the pace of humanoid robot development and production”, Oct 8 ↩
- Tesla News Wire summary, “Optimus Output Reportedly Jumps to Hundreds a Week, but the Robots Still Can't Generalize, and Tesla Hasn't Confirmed a Number”, Oct 10 ↩
- RobCo, “The next frontier for AI is physical: RobCo becomes a unicorn and offers employees liquidity”, Oct 5 ↩
- EU-Startups, “Munich-based RobCo becomes a robotics unicorn as valuation doubles in nine months” ↩
- TechCrunch, “China's Manus raises over $500M in first funding round since split with Meta”, Oct 8 ↩
- TechCrunch, “Popular AI leaderboard Arena nearly doubles valuation to $3.1B valuation in 10 months”, Oct 8 ↩
- TechNode, “Sports-training robot maker PONGBOT raises hundreds of millions of RMB in Series B”, Oct 12 ↩
- Reflection, “Introducing Beam: Reflection’s 501B open-weight model” ↩
- Hugging Face, “IQuestLab/IQuest-Q1 model card” ↩
- Microsoft, “Microsoft-Decision-1: Our model for fast decision-making” ↩
- Hugging Face Papers, “MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement” ↩
- Anthropic, “Introducing the Anthropic Cyber Mission” ↩
- Anthropic, “Building on our commitment to American scientific discovery” ↩
- arXiv, “Long-WAM: Scaling the Context of World-Action Models” ↩
- GitHub, “NVlabs/LongLive” ↩
- arXiv, “SuperNav: An Agentic Navigation System for Any Task in Any Scene” ↩
- GitHub, “zju3dv/SuperNav” ↩
- arXiv, “DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training” ↩
- arXiv, “In-context Robot Learning Made Simple: A Democratized Recipe for Manipulation Tasks” ↩
- Robotics map ↩