Five Ages Series A

Zhongke Diwiji (Five Ages) Deep Dive: A CAS-Institute-of-Automation and Tsinghua Spin-out Whose FAM Embodied Foundation Model Uses Action Optical Flow for Cross-Embodiment Generalisation, Reaching 97% Success from 3 to 5 Demos, with EraBrain and Jiling Robots Serving Sinopec, Leapmotor and Midea

Zhongke Diwiji (Five Ages) was founded in September 2024 by the CAS Institute of Automation with the Tsinghua Sun Fuchun team, led by CEO Liu Nianfeng; its FAM embodied foundation model uses action optical flow for cross-embodiment generalisation, reaching 97% success from 3 to 5 demos, and its EraBrain and Jiling-L1/S1 robots already serve Sinopec, Leapmotor and Midea, after A1/A2 rounds totalling over RMB 1 billion at roughly RMB 7 billion valuation.

Executive Summary

Zhongke Diwiji (Five Ages, full legal name Zhongke Diwiji (Hangzhou) Intelligent Technology Co., Ltd.) is an embodied-AI company founded in September 2024 by the CAS Institute of Automation together with the Tsinghua team of Professor Sun Fuchun. The technical core is the in-house FAM embodied foundation model family (iterated to FAM 2.0 in September 2026): an action-optical-flow representation unifies the motions of different robot bodies into one shared language, which an embodiment-configuration encoder translates into per-robot commands for cross-body reuse; a heatmap representation that preserves three-dimensional geometry enables few-shot learning — 3 to 5 real demonstrations per task reach a 97 percent basic-task success rate, with data requirements about 1 percent of the conventional approach. The product line pairs the EraBrain embodiment module with two bodies: the wheeled dual-arm Jiling-L1 and the column-type Jiling-S1 unveiled at the October 2026 CIIF industrial fair. Customers include Sinopec, Leapmotor and Midea among ten-plus accounts, and in 2026 the company won the ARNOLD challenge at the CVPR Embodied AI Workshop, ahead of teams from NVIDIA, UC Berkeley and Stanford. Within under two years it completed six funding steps from seed through A1/A2, with A1/A2 totalling over RMB 1 billion and a valuation of roughly RMB 7 billion (single-source estimate).

Caveats: the company is private with no audited financials; the 97 and 96 percent success figures are company-reported with no public evaluation protocol; round naming and lead investors conflict across sources, adjudicated item by item in the Discrepancies section.

The verdict: the clearest technical wedge in this cohort of embodied-AI companies, validated deployment-first. Action optical flow attacks the industry's most expensive cost structure — one model per robot — and the multi-robot relay demonstration (three machines completing depalletizing, pouring and palletizing with no human in the loop) shows the company is aiming at process-level collaboration, the key leap from demo to production.

Company Overview

The company was founded in September 2024 in Hangzhou. The founding lineage is explicit: founder and CEO Liu Nianfeng holds a PhD from the CAS Institute of Automation under academician Tan Tieniu; co-founder and chief scientist Sun Fuchun is a tenured professor at Tsinghua's Department of Computer Science, an IEEE, CAA and CAAI Fellow and a National Distinguished Young Scholar, with long-standing work in intelligent control, robot navigation and dexterous manipulation; co-founder and young chief scientist Huang Yan, also from Tan's group, was an assistant researcher (2017 to 2019) then associate researcher (2019 to 2026) at the CAS Institute of Automation and a full researcher from 2026, working on VLA, VLN and multimodal intelligence. Other co-founders include algorithm director Liu Jing, AI director Li Bingbing and R&D director Wang Qingbin.

The founding thesis targets the industry's costliest pain point: one model per robot. Conventional practice collects thousands of demonstrations and re-trains for months for every new body or line. Liu's judgment is that the decisive variable is not body form but brain generalization — robots fail not because they cannot act, but because learning to act costs too much. The company therefore deliberately narrowed its position: no performative humanoid showpieces, but a reusable, low-cost-adaptation embodiment brain.

The pace has been unusually fast: founded September 2024; seed and seed-plus rounds in February and March 2025; Pre-A and Pre-A+ within one month in February 2026 (Sequoia China leading the Pre-A); an A round in May; A1 and A2 in August 2026 totalling over RMB 1 billion; FAM 2.0 released in September with back-to-back first places in the WorldArena benchmark; and a CIIF debut in October that became one of the fair's focal points with a three-robot relay demonstration and two new products.

Technology

Action Optical Flow: The Cross-Embodiment Mechanism

FAM's key choice is to avoid regressing joint commands for a specific robot. Instead, motion sequences from different bodies are encoded into a shared optical-flow-style action representation that learns the motion regularities of grasp, move and place themselves, forming a transferable common action language. The embodiment-configuration encoder reads body information — joint counts, degree-of-freedom layout, link dimensions — and translates the common language into executable per-robot commands. In the team's own phrasing: the model learns how to do a task, not how a particular robot does it. The same capability thus transfers across arms, humanoids and wheeled platforms.

This architectural idea parallels Hydra-0, the framework from a joint NVIDIA-Harvard team that conditions world modeling and control on action flows and cuts robot motion error by 90.4 percent. The company's claim is to have industrialized the idea first — via FAM 2.0 in September 2026 — ahead of proposing it.

Few-Shot: Heatmap Representation and Data Efficiency

Few-shot capability comes from the heatmap technique: the action representation preserves and exploits three-dimensional spatial geometry, raising the utility of scarce demonstration data. The FAM 2.0 architecture diagram shown at CIIF divides the work across four components — action-flow encoding (action representation), ego-data pretraining (video encoding) and heatmap representation (spatial geometry). The deployment claim: 3 to 5 real demonstrations per task reach a 97 percent basic-task success rate, with data needs around 1 percent of conventional methods. In a factory these two numbers directly set deployment cost and schedule — the conventional route implies thousands of demonstrations and months of retraining per line.

The World-Model Layer and Human-in-the-Loop

An embodied world model and the VLTA multimodal robot model support planning and scene adaptation; human-in-the-loop reinforcement learning refines behavior with limited real data. Public detail on this layer is thinnest — model scale, training-data composition and the sim-to-real mix are all undisclosed.

Third-Party Validation: ARNOLD and Two Self-Reported Numbers

The company's strongest external signal is first place in the ARNOLD challenge at the CVPR 2026 Embodied AI Workshop, ahead of NVIDIA, UC Berkeley and Stanford entries, followed by consecutive first places in the WorldArena benchmark in September 2026. By contrast, the 97 percent few-shot figure and the substation-inspection figure (over 96 percent success, 70 percent reduction in on-site human task frequency) are company-reported without a public protocol. Diligence should run an independent benchmark across bodies on held-out tasks.

Product and Commercial Model

Product Matrix: A Brain and Two Bodies

EraBrain is the core product: a sellable embodiment-brain module that also powers the company's own bodies. The wheeled dual-arm Jiling-L1 targets mobile operation and multi-robot collaboration; the column-type Jiling-S1, unveiled in October 2026, targets fixed-station precision assembly and warehouse picking — height 161 to 189 cm, 5 kg single-arm rated load, 19 degrees of freedom excluding the end effector, column lift plus 360-degree waist rotation for a large working volume, modular quick-change tooling for flexible changeover, and high-precision force control for fine work. The two bodies extend the brain's application coverage in complementary directions.

Multi-Robot Collaboration: From Single-Machine Demos to Process Work

At the October 2026 CIIF (26th edition, National Exhibition and Convention Center Shanghai, over 3,200 exhibitors), three Jiling-L1 units ran a complete material-flow relay: unit one depalletizes and delivers boxes to a transfer station; unit two picks, pours and returns empties; unit three takes empties, collects material at the pouring station and carries it to the palletizing zone — with no human intervention. The engineering difficulty is timing: handoff windows are narrow, and one robot's delay propagates down the chain. The demonstration matters because it shows a process — part of a broader 2026 trend of group intelligence entering factories alongside the Beijing Humanoid Robot Innovation Center and Zeekr's smart-factory training runs.

Customers and Scenarios

Domestically the company serves Sinopec, Leapmotor and Midea among ten-plus core customers across logistics, electronics, auto parts, light manufacturing and energy operations. The energy-side metric: a robot carrying the embodied operation brain completed indoor-outdoor high-precision inspection at a substation with over 96 percent operational success and a 70 percent reduction in on-site human task frequency. Overseas progress is concrete: a real-scene validation of general tote handling at the headquarters of a global manufacturer in Germany — randomly placed, mixed-size totes handled continuously without human intervention — plus overseas orders worth hundreds of millions of RMB now in delivery, with Europe, Australia-New Zealand, Japan and Korea planned within six months. Liu's framing: Europe's labor-structure shift and a new automation upgrade cycle create an opening for Chinese embodied-AI companies.

Business Model

Software plus hardware plus industry solutions: the brain module is sold standalone and inside the company's own robots, delivered as configurable robots with a general brain. The model carries high ticket sizes and controls deployment quality, but it is delivery-heavy, hardware-dragged on margin, and in tension with the pure brain-licensing route (licensing to third-party body makers) that would be the end-state of the brain thesis — no public third-body partnership exists yet.

Competitive Landscape

The embodied foundation-model segment is crowded in 2026: Skild AI and Physical Intelligence lead on capital abroad; Zhiyuan, Galbot, Xingdong Jiyuan and domestic peers compete for foundation-model positions; national teams such as the Beijing Humanoid Robot Innovation Center push multi-body collaboration. The right comparison axes are generalization, sample efficiency and deployment reliability rather than stagecraft.

XVERSE's differentiation is threefold. First, the action-optical-flow plus few-shot wedge attacks the one-model-per-robot cost structure directly, backed by the ARNOLD title. Second, the CAS-Institute-of-Automation-plus-Tsinghua scientist roster provides academic resources and a talent pipeline. Third, reference customers span energy (Sinopec), automotive (Leapmotor) and appliances (Midea), spreading exposure better than single-scene rivals. The weaknesses are equally clear: a two-year track record, no independent benchmark protocol, funding an order of magnitude behind the best-capitalized rivals, and the substitution path where hardware makers internalize the brain — if the brain fails to establish an independent ecosystem position, its value accrues to the body makers.

Financing and Capital Structure

At least seven capital events are on record: a seed round on February 19, 2025 (Zhuoyuan Asia); a seed-plus round of tens of millions of RMB on March 17, 2025 (Lihe Kechuang, Lihe Capital, Tongchuang Weiye, Zhuoyuan Asia, Yuejiang Technology); Pre-A and Pre-A+ within one month in February 2026 (Sequoia China leading Pre-A); an A round of hundreds of millions of RMB on May 27, 2026 (Shanghai Semiconductor Industry Investment, Hongruida, Futeng Capital, Guoke Capital, China Venture Capital, with media variants naming Futeng as lead and Sequoia participating); and consecutive A1/A2 rounds in August 2026 totalling over RMB 1 billion (A1: Xigao Investment, BOC AIC funds, Sanfeng Investment; A2: Xineng Venture Capital, BOC AIC funds, Zhongshan Venture Capital, Hongruida, Beyondsoft and a Zhejiang NEV sub-fund). A third-party database aggregates disclosed rounds at about US$201.6 million, with the latest valuation around RMB 7 billion (August 2026, single source).

The investor mix blends state capital with market funds: BOC AIC funds, Xigao Investment and the Zhejiang NEV sub-fund provide state-industrial alignment, while Sequoia China's continued participation signals market confidence. The combination brings both capital and scenario access — winning an energy customer like Sinopec is unlikely to be unrelated to that background.

DISCREPANCIES

1. Round naming and timing conflict. Source A (third-party database): seed round February 19, 2025 (Zhuoyuan Asia) and seed-plus March 17, 2025 (Lihe Kechuang et al.). Source B (media): Pre-A and Pre-A+ completed within one month in February 2026, Pre-A led by Sequoia China. The two may describe the same events under different labels (seed-plus versus Pre-A), but a one-year gap cannot be reconciled, and the database may have missed the Pre-A rounds. Adjudication: record both; defer to corporate registry filings. Source grade: DB versus MEDIA conflict.

2. A-round participants. Source A (database): Shanghai Semiconductor Industry Investment, Hongruida, Futeng Capital, Guoke Capital, China Venture Capital. Source B (media): Futeng leading with Sequoia China participating; an earlier master table also lists Pudong Sci-Tech Investment and Zhuoyuan Asia. Adjudication: database as the skeleton, media variants flagged; the lead is ambiguous. Source grade: DB versus MEDIA conflict.

3. The RMB 7 billion valuation. A single database source (August 2026 basis), no company disclosure. Adjudication: mark as estimate, not a disclosed valuation. Source grade: DB.

4. The two success rates' applicability. The 97 percent is a laboratory figure for basic tasks from 3 to 5 demonstrations; the over-96 percent is a field substation-inspection figure (with a 70 percent cut in human task frequency). Different tasks, different settings, not interchangeable; both company-reported. Adjudication: list separately with flags. Source grade: OFFICIAL (self-reported).

5. Invention rights for action optical flow. The company claims first-of-kind action-optical-flow technology, but the architecture closely parallels Hydra-0 (NVIDIA-Harvard et al., 2025, 90.4 percent motion-error reduction); the company itself claims first industrial productization, ahead of proposing. Adjudication: standardize on first industrial productization. Source grade: OFFICIAL and MEDIA.

6. The weight of the ARNOLD title. The database and the company's CIIF exhibit both cite the 2026 CVPR Embodied AI Workshop ARNOLD challenge win over NVIDIA, Berkeley and Stanford. An academic challenge is not the same dimension as industrial deployment performance. Adjudication: record as the strongest third-party signal, context-limited. Source grade: MEDIA and OFFICIAL.

7. Overseas orders versus plans. Hundreds of millions of RMB in overseas orders in delivery coexists with a plan to cover Europe, Australia-New Zealand, Japan and Korea within six months; the former happened, the latter is planned. Some reports conflate the two into completed overseas layout. Adjudication: separate fact from plan. Source grade: OFFICIAL.

8. Legal name versus brand. The legal name is Zhongke Diwiji (Hangzhou) Intelligent Technology Co., Ltd., the English brand Five Ages; reports also use shorthand forms, and the similarly named but unrelated Zhongke Guiji invites confusion. Adjudication: use the legal name with Five Ages in parallel. Source grade: OFFICIAL.

9. Fourth-batch A1/A2 investors and registry data. Zero2IPO (Qingke) disclosed on 2026-10-30 additional A1/A2 investors including BOC Capital, Puke Investment and Oriental Fortune, supplementing the report’s existing A1/A2 list (BOC AIC funds, Xineng Venture Capital et al.). Enterprise registry records additionally show registered capital of CNY 2.327243 million, 2025 social-insurance participation of 7 people, and national-level "Nazha Enterprise" plus provincial "Future Unicorn" designations, which sit on a different disclosure dimension from the reported billion-scale financing. Adjudication: record the registry and Zero2IPO disclosures as supplements; the financing-scale figure remains as reported. Source grade: DB.

Risks and Diligence Gaps

First, independent validation of the success rates is the longest gap. Both figures lack a public protocol; diligence should benchmark across bodies on held-out tasks against per-body baselines — if the advantage is not significant, the technical wedge fails.

Second, scaled reliability is unproven. The relay demo and the substation deployment are a demonstration and an early installation; round-the-clock reliability, integration with existing line IT and MES, and total cost of ownership at scale are all unverified. Track reference-deployment renewals and expansion orders.

Third, financing transparency. Round labels are tangled across seed, seed-plus and Pre-A, the A-round lead is ambiguous, and the valuation rests on one source. Verification requires registry filings and shareholder agreements.

Fourth, a crowded field with richer rivals. Embodied foundation models drew the densest capital of 2025 and 2026; the company must convert the ARNOLD title into reproducible customer-side benchmarks and defend its position with customer lock-in rather than narrative.

Fifth, overseas execution. The orders are a strong signal, but a three-continent rollout within six months demands localization and support capacity beyond the company's current scale; execution risk is real.

Sixth, the brain's ecosystem position. The brain-plus-own-bodies model sits in tension with the pure brain-licensing end-state; if third-party body makers internalize the brain, the addressable market compresses.

Seventh, the diligence checklist. Independent benchmark results, MTBF and cycle-time data from live deployments, line-of-business revenue and margins, overseas delivery terms, and the valuation basis all need priority verification.

Composite Assessment

Technology: 7.5. Action optical flow plus heatmap few-shot is a clear, novel wedge, third-party-corroborated by the ARNOLD title and WorldArena firsts, with checkable FAM 2.0 components (flow encoding, ego-data pretraining, heatmap representation). Deductions: two self-reported success rates and a thin world-model layer.

Business: 6.5. Ten-plus core customers across energy, automotive and appliances, a live substation deployment, a German validation and overseas orders in delivery. Deductions: undisclosed revenue and margins, unproven scaled reliability, an aggressive expansion plan.

Financing: 7. Six funding steps in under two years, A1/A2 above RMB 1 billion, BOC AIC and Sequoia China on the register — first-tier-adjacent capital adequacy. Deductions: tangled round records and a single-source valuation.

Product: 7. A complete brain-plus-two-bodies matrix, Jiling-S1 specified for industrial work (19 DoF, 5 kg, modular tooling), and a rare process-level relay demonstration at CIIF. Deductions: short production record for new hardware and no third-body ecosystem yet.

Market: 7. Embodied AI sits at the intersection of capital and policy tailwinds, manufacturing labor shifts are a long-run positive, and Europe offers increment. Deductions: fierce competition, a valuation embedding high expectations, and the structural risk of brain internalization by body makers.

Composite: 7.0. Zhongke Diwiji is the most sharply focused thesis in this batch of embodied-AI companies: it has staked everything on cross-embodiment generalization and few-shot learning — the single variable that sets deployment cost — and has supported that thesis with the ARNOLD title, a process-level relay demo and reference customers to the limit of what a two-year record allows. The physical-AI verdict: the meeting point between embodiment brains and industrial simulation is the physical feasibility of motion — if the motion regularities learned by action optical flow are to transfer to new bodies and scenes, physical-constraint checking is ultimately unavoidable; the company's depth of simulation usage is undisclosed, and that is the one unclosed loop in the technical narrative.

References

  1. Tencent News CIIF walkthrough (2026-10-18): the three-robot Jiling-L1 material-flow relay with no human intervention; the EraBrain launch targeting cross-embodiment generalization and few-shot learning; Jiling-S1 specifications (height 161 to 189 cm, 5 kg single-arm load, 19 DoF excluding end effector, column lift plus 360-degree waist rotation, modular quick-change); target sectors of electronics, auto parts and light manufacturing
  2. Robot Lecture Hall CIIF deep-dive (2026-10): timing-difficulty analysis of multi-robot collaboration (narrow handoff windows, error propagation); the FAM 2.0 architecture diagram (action-flow encoding, ego-data pretraining, video encoding, heatmap representation); the Hydra-0 parallel (90.4 percent motion-error reduction) and the first-industrial-productization claim; the division of labor between Jiling-S1 and Jiling-L1; Liu Nianfeng's judgment on Europe's labor-structure shift
  3. Toutiao Caijing Zaocan (2026-10-24): the company just past two years old; A1/A2 above RMB 1 billion (BOC AIC funds, Xineng Venture Capital et al.); Sequoia China leading the Pre-A; substation inspection above 96 percent success with a 70 percent cut in human task frequency; tote-handling validation at a global manufacturer's German headquarters and hundreds of millions of RMB in overseas orders entering delivery
  4. RoboRadar company file (accessed 2026-10): the funding table (seed February 19, 2025, Zhuoyuan Asia; seed-plus March 17, 2025, tens of millions, Lihe Kechuang, Lihe Capital, Tongchuang Weiye, Zhuoyuan Asia, Yuejiang; A round May 27, 2026, hundreds of millions, Shanghai Semiconductor Industry Investment, Hongruida, Futeng, Guoke, China Venture Capital; A1/A2 August 2026 above RMB 1 billion with the full investor list); the latest valuation of RMB 7 billion (August 2026); cumulative disclosed funding about US$201.6 million; the CVPR 2026 ARNOLD challenge win over NVIDIA, Berkeley and Stanford; the September 21, 2026 FAM 2.0 release with consecutive WorldArena firsts
  5. Longrun portfolio monthly letter: founded September 2024 by the CAS Institute of Automation with Tsinghua's Sun Fuchun lab; Liu Nianfeng as CEO, Sun Fuchun as chief scientist, Huang Yan as young chief scientist; the FAM and VLTA models; Sinopec, Leapmotor and Midea customers plus the German validation
  6. Team records (RoboRadar, 2026-10): Liu Nianfeng (PhD, CAS Institute of Automation, under Tan Tieniu); Sun Fuchun (Tsinghua tenured professor, IEEE, CAA and CAAI Fellow, National Distinguished Young Scholar); Huang Yan (CAS Institute of Automation 2017 to 2026, full researcher from 2026, VLA, VLN and multimodal focus); co-founders Liu Jing (algorithm director), Li Bingbing (AI director), Wang Qingbin (R&D director)
  7. Technical background (public): the Hydra-0 framework (joint NVIDIA-Harvard team) conditions world modeling and control on action flows, cutting robot motion error by 90.4 percent
  8. Technical background (public): conventional embodied models need thousands to tens of thousands of demonstrations and months of retraining per new task; 3-to-5 demonstrations and a 1 percent data requirement directly reset deployment cost and schedule
  9. Technical background (public): few-shot learning depends on preserving three-dimensional spatial geometry (heatmap representation); human-in-the-loop reinforcement learning refines behavior with limited real data
  10. Industry background (public): multi-robot collaboration moving from lab to factory in 2026 — the Beijing Humanoid Robot Innovation Center's multi-body, multi-scene, multi-task autonomous collaboration and Zeekr's 5G smart-factory multi-robot training
  11. Industry background (public): the 26th CIIF (October 12, 2026, National Exhibition and Convention Center Shanghai, 310,000 square meters, over 3,200 exhibitors, 24 countries and regions), 56,000 square meters of robot halls, and the sector's turn from performance to work
  12. Commercial background (public): Europe's manufacturing labor-structure shift and automation upgrade as a market opening for Chinese embodied-AI companies; the plan to cover Europe, Australia-New Zealand, Japan and Korea within six months
  1. Zero2IPO Research (2026-10-30): disclosed fourth-batch A1/A2 investors including BOC Capital, Puke Investment and Oriental Fortune, a supplemental caliber versus the report’s existing list; the RoboHub financing table differs from the database on angel, angel-plus and Pre-A-plus naming.
  2. Enterprise registry (queried 2026): registered capital of CNY 2.327243 million, 2025 social-insurance participation of 7 people; national-level "Nazha Enterprise" and provincial "Future Unicorn" designations, a different caliber from disclosed financing scale.
This report is compiled from public sources and does not constitute investment advice. The company is private with no audited financials; both success-rate figures are company-reported without a public evaluation protocol, funding rounds conflict in naming across sources, and the valuation rests on a single-source estimate; uncertainties are flagged throughout.

Related reports66

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.