XIRRA Deep Dive: A Tsinghua-Bred Team with an NTU Scientist Building the LPM Large Physics Model that Unifies World Model and Policy via Bellman Optimality, a 2026 Embodied-Physical-AI Challenger
XIRRA was founded in Shenzhen in July 2026 by Tsinghua-bred Li Yin and NTU scientist Zhang Hanwang, and builds the LPM Large Physics Model that unifies world model and policy via Bellman optimality; it has raised above CNY 100 million at a valuation near USD 500 million with no disclosed revenue.
Executive Summary
XIRRA (Shenzhen Xirangkaiwu Technology Co., Ltd.) is an embodied physical-AI startup founded in Shenzhen in July 2026 by Li Yin, a Tsinghua alumnus and former chief technology officer of Huawei Cloud's large-model organisation, together with co-founder and chief scientist Zhang Hanwang, a professor at Nanyang Technological University and former Huawei chief scientist for multimodal AI. Its flagship is the LPM, a Large Physics Model whose thesis is that world model and control policy can be unified through the optimal solution of a Bellman equation, organised across L0, L1 and L2 levels. The company has also built XR data, training and evaluation platforms, and reports a number-one trajectory-accuracy result in Track 1 and a top-three physical-adherence result in WorldArena 2.0. It has raised above CNY 100 million at a valuation reported near USD 500 million, with no disclosed revenue.
Company Overview
XIRRA was incorporated in Shenzhen in July 2026. The leadership pairs an operator with a research lead: Li Yin brings a Tsinghua background and former Huawei Cloud large-model CTO experience, while Zhang Hanwang brings an NTU professorship and a Huawei multimodal-chief-scientist record, with a research profile in causal machine learning. The investor base is led by Dunhong Asset, with Huakong Fund, Sanhua Holdings and Yinxinggu Capital among the participants. The company is privately held at the seed stage.
Technology
The central technical claim is the LPM, a Large Physics Model built on the idea that a world model and an action policy need not be two separate modules but can be derived together from the optimal solution of a Bellman equation, so that physical consistency and decision-making share one representation. The company organises the model across L0, L1 and L2 levels, which it describes as ranging from low-level physical priors to higher-level planning and policy, though the precise boundary of each level has not been specified in public materials.
The accompanying XR platforms cover data, training and evaluation. The data platform is intended to generate and curate physical interaction data, the training platform to fit the LPM, and the evaluation platform to score physical adherence and trajectory accuracy. The WorldArena 2.0 results, a first-place trajectory-accuracy finish in Track 1 and a top-three physical-adherence result, are the main public evidence the company cites, though a single benchmark is a weak basis for assessing a research programme.
Product and Commercial Model
XIRRA's product surface today is a research-and-platform stack rather than a shipped commercial product. The LPM and the XR data, training and evaluation platforms are positioned for embodied-AI developers and for robotics companies that need physically consistent world models and policies. No pricing, no paying customer and no revenue have been disclosed. The commercial logic is to ride the embodied-AI capital wave by offering model and infrastructure rather than robots.
Competitive Landscape
XIRRA sits in the same broad field as Physical Intelligence, Skild AI, AMI Labs and World Labs, all of which argue that a learned world model is the substrate for physical intelligence. Its differentiation is the Bellman-optimal-unification framing and the causal-ML research lineage of its chief scientist. The risk is that the field is exceptionally crowded and well funded, and that a July-2026 entrant must show results beyond a single benchmark to be taken as a serious contender.
Financing and Capital Structure
The company raised above CNY 100 million across seed and angel tranches at a valuation reported near USD 500 million, led by Dunhong Asset with Huakong Fund, Sanhua Holdings and Yinxinggu Capital participating. A valuation near USD 500 million for a company founded in mid-2026 is at the upper end for the seed stage and signals strong investor conviction in the embodied-physical-AI thesis, but also prices in substantial execution risk ahead of any product revenue.
DISCREPANCIES
1. Founding date. The company is reported as incorporated in July 2026, while some coverage loosely places the team's formation earlier; the legal date is the reliable anchor. Source grade: OFFICIAL versus MEDIA.
2. Funding and valuation. The round is described as above CNY 100 million at a valuation near USD 500 million; the pre- or post-money basis and the exact amount are not disclosed. Source grade: MEDIA.
3. Zhang Hanwang profile. The chief scientist's citation metrics, including a stated H-index of 81 and 228 papers, are self-reported in coverage and have not been independently verified. Source grade: MEDIA.
4. Benchmark standing. The WorldArena 2.0 results are cited as Track 1 trajectory accuracy first place and physical adherence top three; a single benchmark is the only public evidence and its comparability is limited. Source grade: OFFICIAL, single benchmark.
5. Product form. The LPM is described both as a model and as a platform; the packaging and delivery model are not yet fixed. Source grade: MEDIA.
6. L0-L1-L2 definition. The three-level organisation of the LPM is described without a public, precise boundary between levels. Source grade: MEDIA.
Risks and Diligence Gaps
The primary gap is the absence of a shipped product and revenue. The company is a research-and-platform play at the seed stage, and its traction rests on a single benchmark and a strong financing round rather than on deployed systems. A second gap is technical verifiability: the Bellman-optimal-unification thesis and the cited metrics require independent reproduction. A third gap is crowding: the world-model field is densely populated by far better-capitalised and longer-operating peers. A fourth gap is team depth beyond the two named founders.
Overall Assessment
Technology scores 7.0. The Bellman-optimal unification of world model and policy is a coherent and distinctive research framing, and the chief scientist's causal-ML background is a real asset. Deducted points: no independent benchmark beyond WorldArena 2.0, levels L0-L1-L2 under-specified, product immaturity.
Commercialisation scores 5.5. A platform play on the embodied-AI wave is reasonable, but there is no pricing, no customer and no revenue. Deducted points: pre-product, undefined go-to-market, no deployment evidence.
Financing scores 8.0. Above CNY 100 million at a near-USD-500-million valuation led by Dunhong Asset is a strong seed outcome and a clear conviction signal. Deducted points: valuation prices in execution risk, pre- or post-money basis undisclosed.
Product maturity scores 5.5. The LPM and XR platforms are described but not shipped as commercial products. Deducted points: research-stage, no customer reference, packaging evolving.
Market position scores 7.5. The embodied-physical-AI theme is large and hot, and the team profile is differentiated, but the field is crowded and the company is the youngest entrant. Deducted points: intense competition, late start, benchmark-dependent credibility.
The composite view is 6.7. XIRRA is a well-funded, research-led bet on a distinctive unification thesis for physical intelligence, with a founder profile that stands out in a crowded field, but with the standard early-stage deficits of no product, no revenue and a single benchmark. The most informative next disclosures will be an independent reproduction of the LPM results, a first commercial deployment, and clarity on which level of the stack the company will monetise first.
References
- XIRRA official materials and coverage, 2026: LPM Large Physics Model unifying world model and policy via Bellman optimality across L0, L1 and L2; XR data, training and evaluation platforms.
- Company founding and investor coverage, 2026: Shenzhen incorporation in July 2026; above CNY 100 million raised at a valuation near USD 500 million; Dunhong Asset leading with Huakong Fund, Sanhua Holdings and Yinxinggu Capital.
- Profiles of Li Yin, 2026: Tsinghua background, former Huawei Cloud large-model CTO.
- Profiles of Zhang Hanwang, 2026: NTU professor, former Huawei multimodal chief scientist, causal machine learning, stated H-index 81 and 228 papers.
- WorldArena 2.0 results coverage, 2026: Track 1 trajectory accuracy first place, physical adherence top three.
- Analysis of the embodied-world-model field, 2026: comparison with Physical Intelligence, Skild AI, AMI Labs and World Labs.
- Coverage of Bellman-equation and optimal-control formulations in learned world models, 2026.
- Startup databases and company filings, 2026: legal-person registration, investor roster and seed-stage status.
This report is compiled from public sources and does not constitute investment advice. The company is private with no audited financials; the funding amount, the near-USD-500-million valuation, the citation metrics of the chief scientist and the WorldArena 2.0 standing are drawn from media coverage with no company-disclosed totals, and the LPM's Bellman-optimal thesis has not been independently reproduced; uncertainties are flagged throughout. Data is current as of 5 November 2026.
Related papers
Related tools
Related reports
- Physical Intelligence Deep Dive: A $2.1B Bet on the GPT Moment for Robotics, With pi0.7 Self-Reporting 60-80 Percent Success on Unseen Tasks and Neither Pricing nor Revenue Disclosed for the Pi API
- Skild AI Deep Dive: A General Robot Foundation Model Incubated at CMU's Robotics Institute, With a USD 1.4B Series C at a USD 14B Valuation, a Conflicting Series B Size, and an ARR Figure That Is a Run-Rate Rather Than Revenue
- World Labs: Fei-Fei Li's Spatial-Intelligence Unicorn, Acquired by AMD for $8.2B
- AMI Labs: A USD 1.03B Seed Bet on World Models, LeCun's Non-Generative Path and Its Blank Scorecard
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.