LiberAI Series A

LiberAI (Jiangxian) Deep Dive: From RDT-1B to the Liber-0 World Action Model — a Tsinghua-spinoff physical-world-model company built on the UMI no-embodiment human-data + world-model paradigm

LiberAI (Beijing Jiangxian Technology) is a physical-world-model company founded in December 2025 by Tsinghua PhD Liu Songming. Its flagship World Action Model Liber-0, trained on over 2.2 million hours of human experience, ranked first on the RoboDojo benchmark. The company pairs Liber-0 with the self-developed LiberSense Glove and the LEO robot under a UMI no-embodiment human-data-plus-world-model paradigm to attack the embodied-data bottleneck. In eight months it closed five funding rounds totalling over RMB 1 billion from Sequoia China, Shunwei, and state-backed investors.

Executive Summary

LiberAI (legal entity Beijing Jiangxian Technology Co., Ltd.) is a physical-world-model company founded on 8 December 2025 and headquartered in Haidian, Beijing. Its stated mission is to build the general "brain" for physical AI: a world model that understands physical laws and transfers human manipulation experience into robot actions. Rather than manufacturing complete robots, the company positions itself as a base-model and data-infrastructure provider for embodied intelligence. The founder, Liu Songming, is a Tsinghua computer-science PhD born in 2000 (one source states 2001) and a recipient of the Tsinghua Special Scholarship; he led the RDT-1B and RDT-2 lineage of manipulation base models before founding the company. This report assembles the public record on team, technology, capital, commercial evidence and competitive position, and rules item by item on the seven groups of conflicting public accounts recorded in DISCREPANCIES. Data is current as of 15 November 2026.

The technical core is Liber-0, a World Action Model released on 4 November 2026 and trained on over 2.2 million hours of human experience (about 2.0 million hours of internet human video plus glove, cross-embodiment, ego and teleoperation data). Liber-0 uses a Mixture-of-Transformers architecture with three expert streams — language, vision and action — joined by multimodal attention, and a shared Physical Encoder that fuses state, embodiment context and noisy actions; it jointly predicts future frames and future actions. The company couples Liber-0 with a self-developed LiberSense Glove (0.3 mm thick, 131 g, 21-keypoint hand pose capture) and the LEO robot platform (20-degree-of-freedom five-finger hand, articulated wrist, adjustable waist). In public zero-shot real-robot tests the model reached 63.5% on pick-and-place, 34.0% on short-horizon non-pick-and-place and 10.5% on long-horizon non-pick-and-place, for an overall 36.0% success rate over 600 trials, versus 8.3% / 5.0% / 1.7% for a GPT-6 Astra policy run directly as the robot controller.

LiberAI is the fastest-funded company in this list's coverage of Chinese physical-AI ventures: five rounds in eight months, cumulative financing reported above RMB 1 billion, from Sequoia China, Shunwei, Meituan Longzhu, ZhenFund, 360, CDB Capital, China Reform, CMC, Binfa, Zhongguancun Science City, Fortune Venture, and Gaoxin Capital. The valuation is reported at about RMB 5 billion by a single source and is treated as unverified. Technically the company is credible and leads the Chinese world-model / embodied-base-model cohort on the RoboDojo benchmark, but it remains pre-revenue, pre-product-pricing, and its long-horizon reliability (10.5%) is early-stage. The signals worth tracking, in order, are an independent fixed-task evaluation, a reproducible multi-scene baseline beyond single-task demos, the first named paying customer, and reconciliation of the valuation.

Company Overview

Beijing Jiangxian Technology Co., Ltd. was registered on 8 December 2025 in Haidian, Beijing, with registered capital of RMB 1,460,400 and Liu Songming as the legal representative. The founding narrative is unusually concentrated on one person: Liu Songming, born 2000 (one source says 2001), was a Tsinghua computer-science PhD student, a recipient of the Tsinghua Special Scholarship (awarded to roughly ten students per year), and ranked first in his cohort; he was advised by Professor Jun Zhu, a scholar in video-generation models, and published first-author work at ICML and NeurIPS. In October 2024, as a Tsinghua doctoral student, Liu was the first author of RDT-1B, a billion-parameter bipedal-robot diffusion base model released about one month before Physical Intelligence's pi0, which combined large-scale multi-robot data pre-training with a Diffusion Transformer and demonstrated zero-shot generalization to unseen objects and few-shot learning of new skills. In 2025 he released RDT-2, which shifted toward no-embodiment (UMI) human data, and in December 2025 he founded LiberAI.

The co-founder, Lin Fanqi, comes from the Tsinghua Institute for Interdisciplinary Information Sciences and was advised by Professor Yang Gao; Lin was a co-author of the paper "Data Scaling Laws" that first proposed the embodied-domain law linking data diversity to generalization, and received an ICLR Oral and a CoRL X-Embodiment Best Paper award. The team is reported at 30 to 50 people, with over 70% from Tsinghua, and has absorbed former engineers from UBTECH, DJI and Tesla Optimus to cover hardware landing. This academic-plus-engineering composition is the company's stated answer to the typical "strong algorithm, weak hardware landing" weakness of university spin-offs. From the public corporate registry the business scope covers technology development, robot-related software and hardware, and data processing, which matches the stated outward positioning of building base models and a data base rather than complete robots; at a stage with no disclosed revenue or product, the team and academic backing are the most verifiable assets today.

Product and Technology

Liber-0 is the flagship World Action Model. Architecturally it is a Mixture-of-Transformers: three expert streams for language, vision and action interact through multimodal self-attention and jointly predict future image frames and future actions, while a shared Physical Encoder integrates current state, embodiment context and noise-injected actions. The model introduces a unified Embodiment representation that encodes human geometry, hand configuration and robot kinematics as context, so the model learns not only what action to take but how different bodies should perform the same task, enabling cross-embodiment skill transfer. Auxiliary tasks — visual question answering, image segmentation, image editing and video understanding — are trained jointly to strengthen the model's grasp of task semantics, spatial relations and operation progress. Pre-training runs in three stages: first, large-scale video learning of future-frame prediction; second, joint learning of visual change and actual action using human video, glove and robot data; third, raising the sampling weight of glove and teleoperation data to sharpen fine manipulation. By the end of stage three the model reaches PnP and Non-PnP average progress scores of 81.4 and 72.3 with no additional real-robot post-training; further post-training adds only 2.2 and 1.5 points.

The hardware stack has two own-developed pieces. The LiberSense Glove is a flexible glove, 0.3 mm thick and 131 g, that records 21-keypoint hand pose and motion trajectory and synchronizes first-person and dual-wrist cameras to capture both the scene a person sees and the action actually performed; crucially it does not require a robot at the collection site, so a person can wear it in kitchens, offices, shops and industrial environments. The LEO robot platform integrates a five-finger dexterous hand, a flexible wrist joint and an adjustable waist, with wrist three-axis rotation, built to approximate the human data-collection embodiment and camera views. A third artifact is an open interactive simulator (sim.liberai.work) where a user uploads an object image, generates a desktop scene, and runs inference to observe Liber-0's behaviour in simulation.

Data Strategy and the "Data Pyramid"

The central problem LiberAI attacks is the embodied-data bottleneck: teleoperation data is slow and expensive, and robot demonstration data does not scale. Its answer is a three-layer "data pyramid." The bottom layer is internet human video (largest scale); the middle layer is UMI glove data (one to two orders of magnitude higher action-supervision precision); the top layer is real-robot teleoperation data (another one to two orders higher quality but most expensive). The design uses the scale of video to backfill scarce physical interaction, and uses modality alignment plus an inductive bias (treating physical causality — force to motion state — as a prior injected into the diffusion architecture) rather than merely guessing what a frame looks like next. The company reports a pipeline that has closed the loop from UMI hardware through data collection, cleaning, and base-model training.

The most cited finding from the Liber-0 technical report is a data-efficiency result: scaling glove data from 1,000 to 10,000 hours raises PnP and Non-PnP average progress by 14.7 and 13.1 points, and 10,000 hours of glove data outperforms 2.0 million hours of internet human video on both benchmarks — a 200x data-volume gap. A second experiment shows that cutting robot demonstrations to 25% of the pool while adding glove data exceeds the 100%-demonstration baseline (75.6 versus 72.8). A third shows that when robot demonstrations cover only pick-and-place, adding human glove non-pick-and-place data lifts the Non-PnP score from 2.2 to 54.6, close to the 56.3 achieved with full robot demonstrations — evidence that high-quality human action data is beginning to transfer entirely new skills to the robot. A fourth, the morphology-scaling experiment, holds glove data volume roughly constant and increases the number of demonstrators from 1 to 100, raising the average progress score from 12.7 to 40.1, suggesting embodiment diversity is itself a scaling axis. The design treats the three questions — where data comes from, at what precision it supervises action, and whether it transfers across embodiments — as one engineering problem solved in three tiers.

Benchmarking and Public Evaluations

LiberAI's public evidence is unusually open for an early company. In the RoboDojo benchmark — a unified simulation-plus-real robot-manipulation suite introduced in 2026 by Berkeley, Tsinghua and Peking University, with 42 simulation and 18 real tasks — the Liber-0 Preview model ranked first (GPT-6 Astra third). On that list Liber-0 Preview scored 30.74 average and 25.52% average success, with long-horizon 45.98, fine manipulation 38.28, memory 37.77, generalization 24.99 and open-semantics 6.68; it ranked first on every sub-item except open-semantics (second). Separately, on 4 November 2026 the company published a GPT-6-generated 60-task zero-shot evaluation: 20 pick-and-place, 20 short-horizon non-pick-and-place, and 20 long-horizon non-pick-and-place, each tested 10 times for 600 real-robot trials, with no robot demonstrations collected for those target tasks. Results were 63.5% / 34.0% / 10.5%, overall 36.0%. Against a GPT-6 Astra policy used directly as the robot controller, the three groups were 8.3% / 5.0% / 1.7%; the company also reports Astra-only averaged 87.3 seconds per single decision and 17.8 minutes per successful task.

Two caveats apply to the zero-shot claim. First, "zero-shot" here means no robot demonstrations for the target task, not the absence of similar human demonstrations, videos, objects or action combinations in training — so the precise novelty (new object instance, new action combination, or new skill category) must be specified. Second, the data-scaling experiments use stage progress scores, which can mask the difficulty of the final step (a robot can approach, grasp and move an object yet fail the final assembly). The company also published an open evaluation gallery (reel.liberai.work) and the browser simulator, which together let outsiders probe the model's true capability boundary rather than view curated demos. It must be stressed that both RoboDojo and the zero-shot evaluation were published by the company or its collaborators, and no independent third-party reproduction has appeared, so the figures should be read as an upper bound on capability rather than a verified baseline.

Business Model, Customers and Go-to-Market

LiberAI states it does not build complete robots and instead provides the underlying model and data base for humanoid and flexible-operation robots. Its stated application directions are humanoid-robot OEMs (exporting the embodied world model and high-quality human-behaviour datasets), industrial flexible equipment (transferring human demonstration to 3C manufacturing and precision assembly), universities and research labs (opening base models and data services), and home-service robots (simplifying programming logic via massive human-behaviour data). Public reporting also places the company in energy scenarios — collaborating with leading energy groups on power inspection and energy-facility operations. As of this report the company has not published product pricing, a free tier, or named paying customers, so its go-to-market is pre-commercial; the commercial proof will be the first named deployment and recurring revenue, which are not yet public.

The company's strategic bet is that human UMI data plus a world model is a more scalable path than pure teleoperation. The LEO platform and the glove are presented as means to that end — the glove lowers collection cost (reported per-collection time cut from about 50 seconds to 10 seconds and cost to about one-fifth of teleoperation, with an hourly rate around RMB 55) and the LEO embodiment approximates the human data-collection condition. Whether this data flywheel actually compounds into a defensible model advantage depends on continued glove-mass-production scale and collection-network expansion, which are not yet independently verified. On the data-service side the company treats its glove collection network and cleaning pipeline as capabilities it could offer externally, but its billing model, data-licensing scope, and whether raw data is opened to customers are all undisclosed, so the verifiability of the business model lags its technology narrative.

Competitive Landscape

LiberAI sits at the intersection of two of the hottest 2026 directions: world models and human-data scaling. Direct overseas analogues are Physical Intelligence (pi0 VLA + diffusion), Generalist (GEN-0), Figure (Helix) and Google DeepMind (DreamerV3); among world-model and spatial-intelligence labs, World Labs (acquired by AMD) and Skild AI (robot foundation model) are adjacent. Domestically the company competes with embodied-base-model and dexterous-hand ventures such as SUDA Tech, XIRRA, Xingji Jiyuan, Xinghai Tu, Kuawei and Tianji, but LiberAI is described as the only one that has run the full UMI-hardware to data-collection to training closed loop while committing to the dexterous hand rather than a gripper. Versus Physical Intelligence the repeatedly cited differentiators are timing (RDT-1B preceded pi0 by about a month; RDT-2 preceded GEN-0 by about a month in adopting the UMI no-embodiment paradigm) and an emphasis on causal alignment rather than frame generation as the primary objective. The principal constraint is that the architecture is not novel end to end — video-action joint modeling (DreamZero), modality-separated MoT, and first-person human pre-training (EgoScale) each pre-exist, so LiberAI's contribution is system integration with its own collection hardware and robot platform, which must be judged separately from algorithmic originality. By paradigm, LiberAI bets on the world-model-plus-human-data route, which contrasts with the end-to-end VLA (vision-language-action) route of Physical Intelligence and Figure: the former predicts future frames and physical causality first and treats action as a byproduct, the latter regresses action directly. The two routes are unconverged and many players are still experimenting in 2026. The defensibility of LiberAI's differentiation is limited — once the UMI paradigm is validated, rivals can quickly add gloves and collection networks, and the true moat will be the scale and unit cost of the data flywheel rather than the architecture itself.

Financing and Capital Structure

LiberAI is privately held and has closed five rounds in eight months, a density rarely seen in early-stage venture. The sequence is: December 2025 seed round (ZhenFund exclusive, about RMB 100 million); March 2026 angel round (Sequoia China first entry, plus ZhenFund); May 2026 angel-plus (about RMB 500 million, Sequoia China, ZhenFund, Shunwei, Meituan Longzhu); June 2026 Pre-A (hundreds of millions, Shunwei led, with Fortune Venture, Yuanhe, Muhua, Sequoia and ZhenFund adding); July 2026 Pre-A+ (hundreds of millions, 360, CDB Capital, China Reform, CMC, Binfa, with Sequoia's fourth consecutive participation); and 17 September 2026 Pre-A++ (hundreds of millions, Zhongguancun Science City, Fortune Venture, Gaoxin Capital, Huaying Capital). Cumulative financing is reported above RMB 1 billion, and one source places the valuation at about RMB 5 billion. Because the company is private, none of the round amounts, the valuation, or revenue is company-disclosed; the figures here are media-reported and the valuation in particular is a single-source estimate treated as unverified (see DISCREPANCIES).

The capital-structure consequence is momentum and expectation: eight months and five rounds from top-tier and state-backed investors signal strong conviction, but they also raise the bar for the next round, which must show reproducible multi-scene baselines rather than single-task demos. For diligence, the items that would carry weight are a company statement on financing and investors, signed customer contracts with values, and disclosed revenue or headcount trajectory; none of these is currently public beyond media summaries. In investor composition the roster spans financial investors (Sequoia China, Shunwei, ZhenFund, CMC, Fortune Venture) and industry or state-backed parties (CDB Capital, China Reform, Zhongguancun Science City, Huaying), a combination rare among early embodied-AI ventures; it supplies ammunition but also brings exit and landing expectations. The eight-month, five-round cadence means existing shareholders keep adding and the valuation is re-marked quickly, which front-loads the milestone bar for later rounds; if the next round cannot replace single-task demos with reproducible multi-scene baselines, the valuation-precursor risk becomes explicit.

DISCREPANCIES

  1. Founder birth year. Account A: most profiles state Liu Songming was born in 2000. Account B: one startup-media profile states 2001. Ruling: record the founder as born around 2000 and flag the exact year as unverified; do not present a single figure as authoritative. Source grade: MEDIA versus MEDIA.
  1. Valuation. Account A: one source reports the valuation has climbed to about RMB 5 billion. Account B: other coverage describes it only as "several times" the initial value without a number. Account C: no company disclosure. Ruling: report cumulative financing above RMB 1 billion as media-reported and label the RMB 5 billion valuation as a single-source estimate that is unverified; do not treat it as a confirmed figure. Source grade: MEDIA versus MEDIA and ABSENT COMPANY DISCLOSURE.
  1. Team size. Account A: the company profile and one database indicate about 30 to 50 people. Account B: an industry radar lists "under 50." Account C: reports of former UBTECH, DJI and Tesla Optimus engineers suggest a small but growing engineering core without a precise headcount. Ruling: report the team as roughly 30 to 50 people, Tsinghua-linked above 70%, and treat the exact number as unverified. Source grade: COMPANY and DATABASE versus DATABASE.
  1. Meaning of "zero-shot." Account A: the company defines zero-shot as no robot demonstrations collected for the target task. Account B: critics note this does not mean the absence of similar human demonstrations, videos, objects or action combinations in training data. Ruling: retain the company's definition but record the boundary — the novelty (new object, new action combination, or new skill category) must be specified before the claim is read as emergent capability; do not present 36.0% as a solved problem. Source grade: COMPANY versus INDEPENDENT ANALYSIS.
  1. Long-horizon reliability. Account A: the company reports 10.5% success on long-horizon non-pick-and-place and highlights that the model begins to complete challenging long-horizon tasks. Account B: independent reading notes that about 90% of those trials did not fully succeed, so 10.5% indicates early-stage capability, not deployable reliability. Ruling: report both the company figure and the independent caveat, and note that progress scores are not full success rates; do not present demos as reliability evidence. Source grade: COMPANY versus INDEPENDENT ANALYSIS.
  1. Technical originality attribution. Account A: the company presents Liber-0's World Action Model and MoT three-stream design as its architecture. Account B: independent analysis points out that video-action joint modeling (DreamZero), modality-separated Mixture-of-Transformers, and first-person human pre-training (EgoScale) each pre-exist, so the systemic integration with own hardware is the defensible contribution, not the individual components. Ruling: record the architecture factually and separate system integration credit from algorithmic originality, rather than ascribing every building block to Liber-0. Source grade: COMPANY and TECHNICAL BLOG versus INDEPENDENT ANALYSIS.
  1. Scope of "human data." Account A: the company states over 2.2 million hours of human experience, of which about 2.0 million hours are internet human video. Account B: critics note that internet video supplies scale while the precise-action supervision that drives fine manipulation comes from glove and teleoperation data, so the 2.2 million figure and the effective training signal are different quantities. Ruling: report the 2.2 million hours as the company's stated total and clarify that action-supervision density, not raw duration, is the relevant efficiency metric; do not let the headline duration stand in for training quality. Source grade: COMPANY versus INDEPENDENT ANALYSIS.

Risks and Due Diligence Gaps

The first gap is commercial immaturity. There is no published pricing, no free tier, and no named paying customer or disclosed revenue; the company is pre-commercial, and its application claims (humanoid OEMs, industrial flexible equipment, energy inspection) are direction statements rather than booked contracts. Diligence should require the first named deployment with results and any recurring-revenue evidence before treating LiberAI as a scaled vendor.

The second gap is long-horizon reliability. The 10.5% success on complex multi-step tasks means roughly nine of ten trials did not fully succeed; progress scores can mask the final-step difficulty. The relevant next metrics are where failure occurs, whether the system detects and recovers, whether human reset is needed, and how many tasks complete per operating hour — none of which is yet published.

The third gap is causal-alignment fragility. The world-model inductive bias (force-to-motion causality as prior) is reported to degrade when contact stiffness changes abruptly or multiple objects couple, exactly the conditions of soft objects, fine contact, and tactile-force control. Position-trajectory similarity does not guarantee correct force, which matters for insertion, tightening, and handling deformable objects.

The fourth gap is the data flywheel's dependence on glove mass production and collection-network scale. The efficiency result rests on 10,000 hours of glove data; whether collection can keep expanding, and whether unit cost (device, calibration, cleaning) makes glove data cheaper than simply growing the video library, is not yet independently verified.

The fifth gap is valuation-precursor risk. Eight months and five rounds have pushed the valuation ahead of revenue validation; the next round must show reproducible multi-scene baselines rather than single-task demos, or the pricing will face a higher bar.

The sixth gap is competition density. The world-model and embodied-base-model field is crowded and architecture-unconverged, with well-funded overseas players (Physical Intelligence, Figure, Google DeepMind) and numerous domestic entrants; LiberAI's differentiation is system integration and the dexterous-hand commitment, which rivals can copy once the paradigm is validated.

Overall Assessment

Technical strength scores 8.0. LiberAI leads the Chinese world-model and embodied-base-model cohort on the RoboDojo benchmark, ships a coherent Mixture-of-Transformers World Action Model, a self-developed glove and the LEO platform, and backs its claims with open zero-shot evaluation and a public simulator. The deduction reflects that the architecture is system integration rather than end-to-end algorithmic novelty, and that long-horizon reliability is early-stage.

Commercial capability scores 6.5. The company has a clear paradigm, top-tier investors, and a defensible data-infrastructure story, but no published pricing, no named customers, and no disclosed revenue; evidence remains direction statements and demos rather than booked contracts.

Financing scores 8.5. Eight months, five rounds, cumulative financing above RMB 1 billion, and a roster from Sequoia China and Shunwei to state-backed funds is exceptional early-stage momentum; the deduction reflects that the RMB 5 billion valuation is a single-source estimate and that the next round must clear a higher commercial bar.

Product maturity scores 7.5. Liber-0 and the LEO platform are released or demonstrated, the simulator is open, and the glove is in scale collection; the deduction reflects that model weights are not public, the release is a Preview, and no commercial product or pricing exists.

Market position scores 8.0. LiberAI is the RoboDojo #1 and the only Chinese entrant that has closed the UMI-to-training loop while committing to the dexterous hand, with ecosystem access far exceeding its headcount; the deduction reflects intense competition and a pre-revenue state.

The composite is 7.6. LiberAI is a credible, fast-moving physical-world-model company whose principal assets are a coherent data-model-hardware closed loop, open evaluation, and exceptional capital backing, and whose principal liabilities are pre-commercial status, long-horizon reliability, causal-alignment fragility, and valuation-precursor risk. The signals worth tracking, in order, are an independent fixed-task evaluation, a reproducible multi-scene baseline beyond single-task demos, the first named paying customer, and reconciliation of the valuation.

References

[1] LiberAI official site, liberai.com — company positioning as a physical-world-model company; Liber-0 release and product loop. [2] Liber-0 technical blog, liberai.com/blog/liber0/ — World Action Model architecture, three-stage pre-training, scaling experiments, zero-shot evaluation protocol. [3] Zero-shot evaluation gallery, reel.liberai.work — real-robot task demonstrations across difficulty tiers. [4] Interactive simulator, sim.liberai.work — browser-based object-upload inference in simulation. [5] Liu S. et al., "RDT-1B: a billion-parameter bipedal-robot diffusion base model," arXiv:2410.07864, 2024 — first author Liu Songming; early large-scale robot-data pre-training with Diffusion Transformer. [6] RDT-2 project page, rdt-robotics.github.io/rdt2/ — UMI no-embodiment human-data paradigm and cross-embodiment zero-shot deployment. [7] RoboDojo benchmark — unified simulation and real-world robot-manipulation suite (Berkeley / Tsinghua / Peking University, 2026); Liber-0 Preview ranked first. [8] Jiqizhixin Pro, "From RDT-1B to Liber-0: How Far Have Base Models Come in Two Years?" (2026-11-05) — detailed account of Liber-0 scaling experiments, architecture and evaluation. [9] Cailianpress / Chinastarmarket Daily (2026-09-17) — Pre-A++ financing and RoboDojo #1 disclosure. [10] 36kr (2026-06) — Pre-A financing and founding background of Liu Songming and RDT lineage. [11] BrandArk (2026-05) — angel-plus round and the UMI plus world-model paradigm explanation. [12] Baidu Baike entries for LiberAI and Liber-0 — consolidated public profile, founding date, model facts and team.

This report is compiled from publicly available information, including the company's papers, official blogs, corporate filings, partner case studies, benchmark results, and media reports. Figures cited under DISCREPANCIES are recorded separately by independent observation and are not merged. This report is for reference only and does not constitute investment advice. Data as of 15 November 2026.

Related tools

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.