Physical AI
for the Home
Confidential · for invited investors and industry partners only.
Investment summary
tunity
- Position: an intelligent growth companion; we build the scarce layer: bodies in homes plus the real-robot data loop
- Seed round · raising ¥100M · 15% equity · about ¥670M post-money
- This ¥100M round is phase 1: one thing in 6 months, the flagship robot on the market; the first feature the RSI platform iterates is inference speed
lights
- Founder: Head of Hardware & Simulation (Shanghai) and World Model Research Leader at Lightwheel; has built the whole embodied-AI chain
- Core competitive edge: our own RSI platform (recursive self-improvement), software and hardware in one loop, iterating and optimizing every robot feature automatically
- Home education: a controllable, paid, reproducible first scenario
- The world model evaluates, real-robot data corrects; prototypes already live, R&D speed is the moat
returns
- 12 months: 20,000 units shipped, hardware revenue ≥ ¥60M, 500K real-robot episodes, into phase 2 with a ¥500M raise
- 24 months: into phase 3, a general-purpose home robot and a household data network in place, target Series A
Why we started: capabilities commoditize, real-robot data stays scarce
Simulation and general models will standardize and get cheap. What stays scarce is a body living in real homes and the real-robot data loop that corrects the world model. That is the gap I saw at Lightwheel.
Simulation and synthetic data are mature; open-source bodies are mature, so a ¥3,000-class robot is buildable; using a world model as a VLA corrector is work I did at Lightwheel. All three line up now.
Add the body and the real-robot data loop, tie them together with the world model, and turn it into an RSI platform — that is our core edge. Start with home education, ship in 6 months, no slideware robots.
Four people, four lanes: technology / commercialization / marketing & fundraising / safety
Lightwheel · Head of Hardware & Simulation (Shanghai) · World Model Research Leader, reporting to founder Chen Xie
Rice Kavraki Lab / RobotΠ Lab · NASA intern (US) · ex-Baidu senior algorithm engineer
Cornell MEng, Electrical & Computer Engineering · admitted to Fudan via competition · Silicon Valley full-stack engineer; 20 years of coding, 10 software copyrights, 4 patents
Employee #2 at Jingya Capital, investing in AI and SaaS; backed unicorns eSign and Zilliz
Co-founder of Vland: tens of millions of RMB from Sequoia and Shunwei, 150 mid-to-large customers, a 30-person R&D team
Took an overseas companion robot from 0 to 1 through engineering and sales: CES 7,000+ sign-ups, nearly 1,000 pre-orders; backed by Li Zexiang’s incubation network
12 years in cloud computing and AI industrialization
Supply chain, fundraising, go-to-market
PhD Mathematics, UC Irvine · MS Computer Science, Rice
Optimization, planning, control · sim-to-real consistency evaluation
Pain points and fixes: it moves, improves, is affordable, and learns from data
| Pain point | What we do |
|---|---|
| Talks, can’t act | A body that really moves; kids teach it with a 30–120 s demonstration or phone teleoperation |
| Shelved after 3 weeks | RSI self-improvement: new skills every month by OTA, better the more it is used |
| Expensive / unsafe | A1 at ¥100-class, Z1 at ¥3,000-class, Q1 under ¥20,000; low speed, light payload, hard stops, task whitelist |
| No household data | Professional data factory + household real-robot loop; parental consent, rewards only for data that enters training |
World model: imagine first, then act, corrected by real outcomes
Hover a module for details.
Three growth companions: a ¥100-class A1 to get in the door, a ¥3,000-class Z1 for volume, Q1 under ¥20,000 for manipulation
Design before features: the product must look designed, and every piece of software and hardware is built to grow up alongside a child. A companion, not a pile of features.
A1
A ¥100-class desktop companion. It talks, sees and shows expressions — no manipulation — so a family can own its first robot at the lowest possible threshold.
Z1
A small bipedal duck for the desk. It walks, sees and shows expressions; kids run balance and gait experiments and teach it motions.
Wheeled dual-arm robot · Q1
Moves and manipulates with two arms. Gen-1 skills: conversation · desk tidying · Gomoku.
A third robot is already in internal tests: a general-purpose home-scenario robot, trialled in home scenes built into our office. It is phase 3.
The RSI platform: software and hardware in one self-improving loop
RSI — recursive self-improvement: train → evaluate → real-robot regression → release → real-robot data back → train again, the whole chain already wired together, with people setting only the goal and the safety limits; within 6 months the first feature, inference speed, iterates fully automatically, while the rest is still launched by hand. This is our core edge.
Software × hardware Core edge
One platform drives both the model pipeline and the robot itself: teleoperation, capture, real-robot regression and monthly OTA all sit inside the loop. Simulation alone, or hardware alone, cannot close it.
Working prototypes Live
The interactive operation platform (teleop · capture · inference) and the SuperAgent five-agent experiment loop are live and wired to Z1 RL.
Compounding Faster each round
Every round adds data and skills; cycle time is the compounding rate. In 6 months, inference speed iterates fully automatically.
inference speed on full auto (6-month target)
auto experiments / week (12-month target)
OTA cadence
Real-robot data loop: data factory + household network
Professional data factory gold standard
- 20 standardized stations, 2,000+ labeled episodes a day (success / failure / takeover)
- The RSI platform decides what to collect: world-model evaluation and real-robot regression find weak spots → collection orders generated automatically
Household network real long tail
- A natural entry point: lessons run every day, so kids’ demonstrations and parents’ corrections come from normal use
- Only in homes: parental consent · on-device de-identification · rewards only for data that enters training
Roadmap: three phases, from the flagship robot to a general-purpose home robot
| Phase | Robot | RSI platform | Skills | Real-robot data | Homes |
|---|---|---|---|---|---|
| Now 2026-09 | Z1 prototype · robot console live; third robot in internal tests in office-area home scenes | SuperAgent prototype: 5-agent loop | — | — | — |
| Phase 1 Months 0–6 · this ¥100M round | Flagship robot launched, first 1,000 units | First feature on full auto: inference speed | 20 | 50K | First customers |
| Phase 2 Months 7–18 · ¥500M raise | ~1.2 m robot with more built-in skills; 20,000 flagship units shipped | Platform mature, covering every skill | 300 total | 3M | 50,000 |
| Phase 3 Months 19–24 | General-purpose home-scenario robot | Automatic cross-body transfer · fully automatic release candidates | 1,000 total | 10M | 100,000 |
Business model: hardware gets in the home, content and trade-ins keep it there
children aged 0–14 in China
students in compulsory education in China
consumer service robots sold worldwide in 2024, mostly floor cleaners
Sources: China NBS and Ministry of Education 2025 bulletins; IFR World Robotics 2025
Three bodies
A1 to get in the door, Z1 for volume, Q1 for manipulation; ≥ ¥60M hardware revenue in 12 months.
Staying in the home
Courses, experiment kits, learning reports; a new generation every six months, upgrade for the difference.
Worth more with use
Institutional edition · SDK; every unit at home keeps feeding the RSI platform.
Benchmarks: Sunday (US) and 未来不远 (China)
Both are betting on the same thing: home robots are won by how fast real-household data comes back. The difference is how that loop runs.
Sunday Robotics Gloves first, robot second
- Product: Memo, wheeled dual-arm, explicitly not a bipedal humanoid; ≈USD 20,000 to hand-build (company-stated), no retail price ever published
- Route: collect data first with a Skill Capture glove built to match its gripper (2,000+ shipped, 500+ paid collectors), then build the robot; the ACT-1 model is claimed to be trained with zero robot data (company-stated)
- Delivery: not purchasable as of 2026-09, a free 50-household invite beta; its funding release says first homes within 2026, its own FAQ states at-scale shipping in 2027–2028
未来不远 Futuring Robot Rent first, sell later
- Product: F1 (2025-09, ~80 kg, 22 DoF); F2 (2026-04, from ¥36,000, no delivery date published)
- Route: ¥3,000 a month, roughly one month per household before rotating on — get the robot into homes first, sell later
- Tech & manufacturing: self-evolving WAM predicts risk, AVLA executes, failures recycle into training; full-stack in-house, 16 hardware generations, unit cost ¥20,000–30,000 (company-stated)
| Dimension | Sunday (US) | 未来不远 (CN) | OriginX AI |
|---|---|---|---|
| Scenario & form | Household chores; wheeled dual-arm, explicitly not bipedal | All-scenario home; wheeled dual-arm | Q1 wheeled dual-arm + Z1 desktop bipedal |
| Product & price | Memo; no retail price, ≈USD 20,000 to hand-build | F1 / F2; F2 from ¥36,000, ¥3,000 a month to rent | A1 ¥100-class; Z1 ¥3,000-class; Q1 under ¥20,000 (all targets) |
| Delivery | Not on sale; free 50-home invite beta; at-scale stated 2027–2028 | Already rented into homes; no F2 delivery date | Prototype; flagship to market in 6 months (target) |
| Route | Gloves collect the data first, the robot comes second | Rent first, sell later: get the robot into the home first | Make the body cheap first, then let the RSI platform make it strong |
| Data source | External glove capture (2,000+) plus paid collectors | Own robots in rented homes | Capture factory + real-robot regression + world-model evaluation feedback; home data once units ship |
| Iteration loop | Glove data into training data; cadence not published | World-model prediction plus failure recycling; cadence not published | RSI platform as one loop: train → evaluate → regression → release → data back (prototype running, not yet at scale) |
| In-house & manufacturing | Robot and glove in-house; manufacturing and capacity not published | Full-stack in-house (motors, joints, torque sensors); 16 hardware generations (company-stated) | Robot and RSI platform in-house; China supply chain |
Their figures come from public reporting and company statements, as of 2026-09; neither publishes a release cadence or says whether real-robot regression is automated. On our side, prices, launch timing and the 20,000 units are targets, not delivered facts; our RSI loop is wired end to end, with inference speed the first function to run fully automatically.
Financing: ¥100M, one thing in 6 months
raise
equity
post-money (approx.)
- Phase 1 · this ¥100M round: flagship robot on the market in 6 months, first 1,000 units; inference speed iterated fully automatically
- Phase 2 · ¥500M raise: RSI platform mature, more skills built into a ~1.2 m robot; 20,000 units shipped in 12 months
- Phase 3: a general-purpose home-scenario robot
Qinzhen Ma · Founder / CEO · qm18@rice.edu · linkedin.com/in/qinzhen-ma ↗