Tashi Zhihang (TARS AI / 它石智航)

China's fastest-funded embodied-AI startup has set a Guinness record and raised nearly $700 million in fourteen months — but has yet to book a single dollar of revenue.

---

How to Read This Report

This report applies four evidence labels consistently throughout. Readers should weight conclusions accordingly.

A choreographed demonstration video is not proof of autonomous production capability. A funding announcement is not proof of revenue. A partnership announcement is not proof of a paying customer. These distinctions are applied without exception below.

---

01Executive Overview

Tashi Zhihang — trading externally as TARS AI, registered in Chinese as 它石智航 — is a Shanghai-based full-stack embodied intelligence company founded on 5 February 2025 ¹². In the fourteen months to April 2026 it raised approximately $697 million across three financing rounds, culminating in a $455 million Pre-A that its backers described as the largest single embodied-AI raise in Chinese history ⁴⁷. Its post-money valuation is reported at CNY 13 billion (approximately $1.88 billion) by two independent market data services ¹¹³, though one commentary source places it closer to CNY 20 billion ($3 billion) ²⁸ — a discrepancy addressed in detail in Section 7.

The company's commercial proposition rests on two pillars. First, a proprietary embodied foundation model called AWE 3.0, which the company describes as a tri-modal system integrating visual perception, language understanding, and motion control ⁶⁹. Second, a physical humanoid robot platform — the A1 — which in March 2026 set a Guinness World Record for sub-millimeter wire harness assembly, completing 105 assemblies in one hour ⁶⁷. Both pillars are currently in industrial pilot and verification stage; the company has not publicly disclosed any revenue, any named paying customer, or any production-scale deployment ²⁹.

The leadership team carries credible pedigree. Chief Executive Chen Yilun comes from Huawei; Chairman Li Zhenyu from Baidu; Chief Scientist Ding Wenchao brings research credentials consistent with the affiliated academic papers discussed in Section 5 ³⁹³⁰. The investor roster is unusually blue-chip for a company of this age: Hillhouse Ventures, Sequoia China, and Meituan Strategic Investment co-led the Pre-A ⁴⁷, and both Beijing and Shanghai state-linked funds participated — reportedly their first-ever direct investment in an embodied-AI company ⁹²⁸.

EDITORIAL INFERENCE: The financing velocity and investor quality signal genuine institutional conviction that humanoid robots will enter industrial manufacturing within a three-to-five-year window, and that Tashi Zhihang has assembled a credible team to compete for that market. What the financing does not prove is that the technology works reliably at production scale, that customers will pay for it at margins that justify the valuation, or that the company can execute hardware manufacturing at volume. Those questions remain open.

The sections that follow examine each of these dimensions in turn: the founding story, the product portfolio, the technology stack, the research output, the video evidence, and the commercial reality as it stands in mid-2026.

Latest news

This module is being compiled — no data to show yet.

---

02The Tashi Zhihang Story

Founding and Context

Tashi Zhihang was incorporated in Shanghai on 5 February 2025 ¹²¹⁶. The founding date places it squarely inside the wave of Chinese embodied-AI startups that emerged in late 2024 and early 2025, catalysed by the commercial traction of Boston Dynamics' Atlas, the rapid iteration of Unitree's G1, and a series of policy signals from Beijing designating humanoid robotics as a strategic industry. The company's name — 它石智航, literally "other stone, intelligent navigation" — is an allusion to the classical Chinese proverb 他山之石，可以攻玉 ("stones from other hills can polish jade"), suggesting an intent to apply techniques from adjacent fields, particularly large language models and computer vision, to the robotics domain.

The Founding Team

The three named principals bring complementary backgrounds that are consistent with a full-stack hardware-software ambition.

Chen Yilun (CEO) is identified across multiple sources as a former Huawei executive ³⁹³⁰. Huawei's relevance here is not incidental: the company has deep expertise in custom silicon, sensor fusion, and large-scale systems integration — precisely the engineering disciplines required to build a vertically integrated humanoid robot. The specific division or role Chen held at Huawei is not publicly disclosed. UNKNOWN.

Li Zhenyu (Chairman) is a former Baidu executive ⁹³⁰. Baidu's Apollo autonomous driving programme gave a generation of Chinese engineers hands-on experience with real-time perception, planning, and control at scale — a skill set that transfers directly to mobile manipulation. Li's specific role at Baidu is not publicly disclosed. UNKNOWN.

Ding Wenchao (Chief Scientist) is named consistently across sources ⁹³⁰. The affiliated research papers discussed in Section 5 — covering tactile manipulation, zero-shot retrieval-augmented generation for robotics, and video-generation-based constraint alignment — are consistent with a research leadership profile, though the precise institutional affiliations of the paper authors are not fully disclosed in the dossier. EDITORIAL INFERENCE: the research output is technically credible and suggests a team with genuine academic depth, not merely a commercial packaging exercise.

The Financing Trajectory

The speed of Tashi Zhihang's fundraising is, by any measure, extraordinary for a company with no revenue.

The angel round closed approximately six weeks after incorporation — an almost unprecedented pace that implies the founding team had investor relationships and a credible pitch deck well before the legal entity existed ¹². The angel+ round, led by Meituan, is strategically significant: Meituan operates one of China's largest logistics and delivery networks, giving it both a potential deployment channel for humanoid robots and a direct commercial interest in the technology's maturation ³.

The Pre-A round, described as oversubscribed ⁷, attracted a notably diverse syndicate. Industrial investors — TCL Ventures, Futeng Capital, Chow Tai Fook Holdings, CATARC Investment — sit alongside financial investors (Hillhouse, Sequoia) and state-linked funds (Beijing Robot Industry Development Fund, Shanghai State Investment Pioneer) ⁹²⁸. The industrial investors are particularly telling: they represent potential customers or supply-chain partners, not merely financial speculators. Whether any of these industrial investors have signed commercial agreements with Tashi Zhihang is not publicly disclosed. UNKNOWN.

The Guinness Record as a Narrative Device

The Guinness World Record set on 10 March 2026 — 105 sub-millimeter wire harness assemblies per hour by the A1 robot — deserves contextualisation beyond its headline value ⁶⁷. Wire harness assembly is one of the most automation-resistant tasks in automotive and electronics manufacturing. It requires fine dexterity, compliance, and the ability to handle flexible, deformable components — properties that have defeated conventional industrial robots for decades. Demonstrating any credible performance on this task, even in a controlled setting, is technically meaningful. The Guinness record format, however, is a marketing instrument as much as a technical benchmark: it establishes a single-session performance figure under conditions controlled by the record-setter, not a sustained production throughput figure validated by an independent customer. The distinction matters enormously for commercial evaluation.

EDITORIAL INFERENCE: The record was almost certainly timed to coincide with the Pre-A fundraising process, providing a concrete technical milestone to anchor investor confidence. This is rational corporate behaviour, not evidence of deception — but it should not be read as proof of production readiness.

---

03Product Portfolio: What Tashi Zhihang Actually Sells

The Honest Starting Point

As of June 2026, Tashi Zhihang does not publicly sell anything. The company is pre-revenue and in industrial pilot and verification stage ²⁹. What it has is a product architecture — a set of hardware and software components that are sufficiently developed to demonstrate, to enter pilot programmes, and to attract the investment described above. The following describes that architecture as reported, with evidence labels applied throughout.

Hardware: The A-Series and T-Series Robots

COMPANY CLAIM: Tashi Zhihang produces two humanoid robot series — the A-series and the T-series — with fully in-house core components and a proprietary sensor array ⁹³⁰.

The A1 is the only specific model named in the available dossier. It is the robot used in the Guinness World Record wire harness assembly demonstration ⁶⁷. Beyond its role in that demonstration and its association with the AWE 3.0 model, detailed hardware specifications — degrees of freedom, payload, reach envelope, joint torque ratings, end-effector design, weight, power consumption — are not publicly disclosed. UNKNOWN.

The distinction between A-series and T-series is not elaborated in any available source. Whether the T-series is a separate form factor (wheeled base, different torso configuration), a different payload class, or a product line at an earlier development stage is not publicly disclosed. UNKNOWN.

The claim of "fully in-house core components" is significant if true. In the humanoid robot industry, core components typically means actuators (motors, gearboxes, or linear actuators), sensors (force-torque, IMU, cameras), and compute hardware. Vertical integration at this level would differentiate Tashi Zhihang from integrators that assemble third-party components, but it also implies substantial capital expenditure on manufacturing tooling and supply chain — costs that are not reflected in any publicly available financial disclosure. EDITORIAL INFERENCE: the "fully in-house" claim is plausible given the funding level but cannot be independently verified from available sources.

Software: The AWE 3.0 Foundation Model

AWE 3.0 is Tashi Zhihang's proprietary embodied foundation model. The company describes it as featuring:

• Tri-modal alignment: simultaneous integration of visual perception, language understanding, and motion control ⁶⁹
• Built-in Failure Recovery: a mechanism described as using latent space future outcome prediction to detect and recover from task failures ⁶⁹
• Zero-shot generalisation: the ability to perform tasks without task-specific fine-tuning, supported by the affiliated research papers ¹⁸²¹

COMPANY CLAIM: Tashi Zhihang describes AWE 3.0 as "the world's first native embodied foundation model with tri-modal alignment" ⁶⁹. This superlative is unverified marketing language. Multiple other companies — including Physical Intelligence (pi0), Figure AI, and Unitree — have released or claimed embodied foundation models with multimodal integration. No independent source corroborates the "world's first" or "native" framing. The capability itself (a unified model spanning vision, language, and action) is technically plausible and consistent with the research output described in Section 5, but primacy cannot be established from available evidence.

The "Failure Recovery via latent space future outcome prediction" claim is technically interesting. Predicting future states in latent space to detect impending failures is a legitimate research direction — related to model-predictive control and world models — but the specific implementation, its reliability in production conditions, and its performance characteristics are not independently documented. UNKNOWN.

Data Infrastructure: SenseHub

COMPANY CLAIM: Tashi Zhihang has accumulated over 100,000 hours of high-quality operational data through its SenseHub suite, which comprises data collection gloves and first-person cameras ⁶⁹.

SenseHub appears to be a teleoperation and data collection infrastructure — the mechanism by which human demonstrations are captured to train the AWE model. This is a standard approach in the field (similar to Apptronik's, Physical Intelligence's, and 1X's data collection pipelines), and the 100,000-hour figure, if accurate, would represent a substantial proprietary dataset. However, the quality, diversity, and task coverage of this data are not independently verified. The figure is reported consistently across two sources ⁶⁹ but both are company-proximate. EDITORIAL INFERENCE: the existence of a structured data collection programme is credible; the 100,000-hour figure should be treated as a company claim pending independent verification.

Dataset: WIYH

COMPANY CLAIM: Tashi Zhihang has open-sourced a dataset called WIYH (World In Your Hands), described as "the world's first large-scale real-world embodied VLTA (Vision-Language-Tactile-Action) multimodal dataset" ⁶. This claim appears in a single source and carries the same "world's first" caveat as the AWE 3.0 model claim. The inclusion of tactile data alongside vision, language, and action is technically distinctive — most open embodied datasets do not include tactile modalities — but the dataset's actual content, scale, and accessibility have not been independently reviewed in the available dossier. UNKNOWN beyond the single-source description.

Product Maturity Summary

Products & versions

A1

Tashi Zhihang's flagship humanoid robot; set a Guinness World Record on March 10, 2026 for the most sub-millimeter wire harness assemblies by a robot in one hour (105 assemblies/hour) powered by the AWE 3.0 model.

AWE 3.0

Proprietary embodied foundation model featuring tri-modal alignment (visual perception, language understanding, motion control) and built-in Failure Recovery via latent-space future-outcome prediction; powers the A1 robot.

---

04Technology Stack: Strengths and the Work That Remains

Architectural Philosophy

Tashi Zhihang's stated approach is full-stack vertical integration: designing its own robot hardware, its own foundation model, its own data collection infrastructure, and its own training datasets. This is the same architectural bet made by Figure AI, Agility Robotics, and — most successfully to date — Boston Dynamics. The logic is that robotics performance is determined by the tightest integration across the entire stack, and that a company relying on third-party actuators, third-party models, and third-party data will always be constrained by interfaces it does not control.

The counterargument is that full-stack integration is extraordinarily capital-intensive, slows time-to-market, and creates organisational complexity that can overwhelm a young company. At $697 million raised in fourteen months, Tashi Zhihang has more capital than most embodied-AI startups to pursue this strategy — but it has not yet demonstrated that it can execute it at production scale.

Perception and Sensing

The AWE 3.0 model's visual perception component is described in terms consistent with the research papers: 3D Gaussian Splatting for scene representation ¹⁸, vision-language model integration for task understanding ²¹, and tactile sensing for contact-rich manipulation ²⁰. This is a technically coherent stack.

3D Gaussian Splatting (3DGS) as a scene representation for manipulation is an active research area. The RobMRAG paper ¹⁸ reports a +7.76 percentage point improvement over a zero-shot baseline using 3DGS-enhanced multimodal retrieval-augmented generation. This is a meaningful but not dramatic improvement, and the baseline comparison matters: zero-shot performance on manipulation tasks is typically low, so a 7.76pp gain over a low baseline may not translate to production-grade reliability.

Tactile sensing is the most technically distinctive element of the stack. The Tac-Man paper ²⁰ describes prior-free tactile-only manipulation of articulated objects achieving near-perfect success rates. If this result generalises beyond the paper's experimental conditions, it addresses one of the hardest problems in contact-rich manufacturing: manipulating objects whose exact geometry and compliance are not known in advance. Wire harness assembly — the company's flagship application — is precisely this type of task. EDITORIAL INFERENCE: the tactile sensing capability, if it performs as described in the paper, is a genuine technical differentiator. The gap between paper results and production performance remains uncharacterised.

Motion Control and Policy Learning

The TapSampling paper ¹⁹ describes an inference-time sampling method using a task-progress-understanding verifier that improves generalist manipulation policies without fine-tuning. This is relevant to the AWE 3.0 model's claimed zero-shot generalisation capability. The approach is policy-agnostic, meaning it could in principle be applied on top of various base models — which suggests the research team is thinking about deployment flexibility, not just benchmark performance.

The EmboAlign paper ²¹ addresses a different problem: using video generation with compositional constraints to provide zero-shot manipulation guidance. The reported +43.3 percentage point improvement over the strongest baseline is a large number, but zero-shot manipulation baselines are typically weak, and the experimental conditions are not independently validated. EDITORIAL INFERENCE: the research direction is sound; the magnitude of the reported improvements should be treated cautiously until replicated.

The Failure Recovery Claim

The built-in Failure Recovery mechanism — described as using latent space future outcome prediction — is the most commercially significant and least documented capability in the stack ⁶⁹. In production manufacturing, a robot that can detect its own failures and recover without human intervention is qualitatively more valuable than one that halts and waits for an operator. The mechanism is described only at a high level in available sources. Whether it has been tested in the industrial pilot environment, what its false-positive and false-negative rates are, and how it performs on the specific failure modes encountered in wire harness assembly are not publicly disclosed. UNKNOWN.

Hardware Integration

The claim of fully in-house core components ⁹ implies that Tashi Zhihang is designing and manufacturing its own actuators, sensors, and possibly compute hardware. This is technically ambitious. The actuator design for a humanoid robot capable of sub-millimeter assembly precision requires either high-resolution proprioceptive sensing (as in MIT's quasi-direct-drive approach) or a combination of compliant actuation and tactile feedback — consistent with the Tac-Man research. The specific actuator architecture is not publicly disclosed. UNKNOWN.

Remaining Technical Challenges

The following challenges are standard for the field and apply to Tashi Zhihang specifically given its stated application domain:

---

05Research, Papers, Authors and Labs

Research Output Overview

Tashi Zhihang, despite being less than eighteen months old, has affiliated research output across four arxiv papers that address distinct sub-problems in embodied manipulation. The papers are technically coherent and address real challenges in the field. The precise institutional affiliations of the authors — whether they are Tashi Zhihang employees, academic collaborators, or both — are not fully disclosed in the available dossier. EDITORIAL INFERENCE: the research is consistent with a team that has genuine academic depth, and the topics map directly onto the company's stated technical priorities.

Paper 1: RobMRAG — Zero-Shot Manipulation via 3DGS-Enhanced MRAG [18]

Core contribution: Integrates 3D Gaussian Splatting scene representations into a multimodal retrieval-augmented generation (MRAG) framework for robotic manipulation, enabling zero-shot task execution by retrieving relevant demonstrations from a memory bank.

Reported result: +7.76 percentage points over a zero-shot baseline.

Editorial assessment: The use of 3DGS for scene representation is technically current — 3DGS has become a preferred representation for real-time novel view synthesis since its introduction in 2023. Applying it to manipulation retrieval is a logical extension. The 7.76pp improvement is modest in absolute terms; the practical significance depends on the absolute performance level of the baseline, which is not reported in the dossier summary.

Paper 2: TapSampling — Inference-Time Sampling with Task-Progress Verifier [19]

Core contribution: A policy-agnostic inference-time sampling method that uses a task-progress-understanding verifier to select higher-quality action sequences at deployment time, without requiring model fine-tuning.

Reported result: Improves generalist manipulation policies without fine-tuning.

Editorial assessment: Inference-time compute scaling for robotics policies is an emerging research direction, analogous to chain-of-thought reasoning in language models. The policy-agnostic framing is commercially attractive — it implies the method could be applied to AWE 3.0 or to third-party base models. The absence of specific quantitative results in the dossier summary limits assessment.

Paper 3: Tac-Man — Tactile-Informed Prior-Free Manipulation [20]

Core contribution: A tactile-only manipulation framework for articulated objects that requires no prior knowledge of object geometry or compliance, achieving near-perfect success rates on a benchmark of articulated object manipulation tasks.

Reported result: Near-perfect success on articulated object manipulation.

Editorial assessment: This is the most directly relevant paper to the wire harness assembly application. Wire harnesses are flexible, deformable, and variable — exactly the conditions where tactile-only feedback is most valuable. "Near-perfect success" in a controlled experimental setting is encouraging but does not establish production reliability. The paper was posted to arxiv in March 2024 ²⁰, predating the company's February 2025 founding — which suggests either that the research team was working on these problems before incorporation, or that the paper is from academic collaborators who subsequently joined the company.

Paper 4: EmboAlign — Video Generation with Compositional Constraints for Zero-Shot Manipulation [21]

Core contribution: A framework that aligns video generation models with compositional constraints to provide zero-shot manipulation guidance, without requiring task-specific data or fine-tuning.

Reported result: +43.3 percentage points over the strongest baseline, zero-shot, data-free.

Editorial assessment: A 43.3pp improvement over the strongest baseline is a striking claim. Zero-shot, data-free manipulation guidance via video generation is a high-value capability if it generalises — it would substantially reduce the data collection burden that currently limits embodied AI deployment. The experimental conditions and baseline definitions are not detailed in the dossier summary, making independent assessment difficult. The result should be treated as a research finding requiring replication, not a production capability.

Research Gaps

The research output covers perception (3DGS-MRAG), policy improvement (TapSampling), tactile manipulation (Tac-Man), and zero-shot guidance (EmboAlign). Notably absent from the public research record: papers on actuator design, robot hardware architecture, manufacturing scalability, or system-level integration. This is consistent with a company that is stronger on the software and AI side than on the hardware manufacturing side — a common profile for embodied-AI startups founded by software and AI engineers.

Company-linked papers

This module is being compiled — no data to show yet.

Authors & labs

This module is being compiled — no data to show yet.

Code & simulation

This module is being compiled — no data to show yet.

Datasets & benchmarks

• WIYH (World In Your Hands)

  Claimed world's first large-scale real-world embodied VLTA (Vision-Language-Tactile-Action) multimodal dataset, open-sourced by Tashi Zhihang.

---

06Media Evidence Library: What the Videos Prove

The Evidence Problem

The video sources in the research dossier ^222324252627 are, without exception, misattributed: they cover a Samsung S26 Ultra teardown, a MacBook teardown, a Framework laptop teardown, a shop vacuum review, an audio subwoofer demo, and a NexPhone hands-on. None of these are Tashi Zhihang content. The dossier's own reconciliation flags this explicitly [dossier summary]. Accordingly, this section cannot draw on the video sources provided.

What can be assessed is the nature of the video evidence that does exist in the public domain based on references in the text sources.

The Guinness World Record Demonstration

The most significant documented video evidence is the Guinness World Record demonstration of 10 March 2026, in which the A1 robot completed 105 sub-millimeter wire harness assemblies in one hour ⁶⁷. Multiple text sources reference this event consistently, and the Guinness World Record organisation's involvement provides a degree of third-party validation that is absent from most robotics demonstrations — Guinness adjudicators are present to verify the count and the conditions.

However, the following questions about the demonstration cannot be answered from available sources:

• What were the exact setup conditions? (pre-positioned components, fixture design, lighting)
• Was the robot operating fully autonomously throughout, or were there human interventions that reset failed attempts?
• What was the failure rate — i.e., how many attempts were made to achieve 105 successful assemblies?
• Was the wire harness variant fixed throughout the hour, or did it vary?
• Was the demonstration filmed and is the footage publicly accessible?

EDITORIAL INFERENCE: The Guinness record establishes that the A1 robot can perform sub-millimeter wire harness assembly at a rate of 105 per hour under controlled demonstration conditions. It does not establish that the robot can do so reliably across the range of harness variants, fixture configurations, and environmental conditions encountered in production manufacturing. The gap between demonstration performance and production performance is the central unanswered question for Tashi Zhihang's commercial case.

What Credible Video Evidence Would Need to Show

For the purposes of evaluating Tashi Zhihang's claims, the following types of video evidence would be editorially meaningful:

The absence of this evidence does not mean the capabilities do not exist. It means they have not been independently documented in a form that meets the evidentiary standard applied in this report.

Media library

Precision Orthopedic Surgical Robot: Tianzhihang Tianji II

Bilibili3.6k views

Tianzhihang Orthopedic Robot 'Tianji' Series - Cervical Spine Surgery

Bilibili1.4k views

Tianzhihang: Orthopedic Robot Leader, Ushering in a New Era of High Growth

Bilibili408 views

---

07Commercial Reality

Revenue Status

Tashi Zhihang is pre-revenue. This is stated or implied consistently across multiple sources ²⁹ and is consistent with the company's age (sixteen months at the time of writing) and its stated position in "industrial pilot and verification stage" ². No source in the dossier identifies a paying customer, a signed commercial contract, or a disclosed revenue figure.

EDITORIAL INFERENCE: Pre-revenue status at this stage is not unusual for deep-tech hardware companies. Boston Dynamics was pre-revenue for most of its first decade. What is unusual is the scale of capital raised against zero revenue — $697 million is a bet on future market creation, not a return on demonstrated commercial traction.

Valuation Conflict

The post-money valuation after the Pre-A round is reported inconsistently across sources:

The CNY 13 billion figure appears in two independent market data services and is the more defensible number. The CNY 20 billion figure appears in a single Chinese-language commentary piece ²⁸ and may reflect a different valuation methodology (e.g., a post-money figure that includes a different share class, or a forward-looking estimate rather than a transaction-implied valuation). The discrepancy — approximately $1.1 billion — is material and unresolved. For the purposes of this report, CNY 13 billion (~$1.88 billion) is used as the base case, with the higher figure noted as a contested alternative.

Investor Composition and What It Signals

The Pre-A investor syndicate is worth examining in detail because it reveals the commercial thesis being underwritten:

The presence of Meituan across two rounds (angel+ and Pre-A co-lead) is the most commercially significant signal ³⁴. Meituan's core business — food delivery, local commerce, logistics — involves enormous last-mile and warehouse operations. If Meituan is investing strategically rather than purely financially, it implies an intent to deploy Tashi Zhihang robots in its own operations. Whether any such deployment agreement exists is not publicly disclosed. UNKNOWN.

CATARC Investment — affiliated with China Automotive Technology and Research Center — signals interest from the automotive supply chain, which is the primary market for wire harness assembly automation ⁹²⁸. This is the most direct line between an investor and the company's flagship demonstrated capability.

The Industrial Pilot Stage

The company describes its wire harness assembly capability as having "entered industrial verification stage" ². This is a specific and meaningful phrase in Chinese manufacturing terminology: it means the technology is being tested in a real factory environment against production specifications, but has not yet been approved for full production deployment. Industrial verification typically involves:

1. Running the system alongside human workers on the same tasks
2. Measuring defect rates, cycle times, and uptime against agreed benchmarks
3. Identifying failure modes and feeding them back to the development team
4. Obtaining sign-off from the customer's quality and engineering teams

The identity of the factory or customer conducting this verification is not publicly disclosed. UNKNOWN. The timeline for completing verification and transitioning to production deployment is not publicly disclosed. UNKNOWN.

Commercial Risks

The following commercial risks are material and not addressed in available public sources:

Unit economics: Wire harness assembly robots must compete with human labour costs in Chinese manufacturing, which — while rising — remain substantially below those in Western markets. The robot's cost per unit, maintenance cost, and productive uptime must combine to deliver a payback period acceptable to industrial customers. None of these figures are publicly available.

Customer concentration: If the industrial verification is occurring at a single customer site, the company's near-term commercial trajectory is highly dependent on that customer's decision. A single rejection could delay revenue by twelve to eighteen months.

Hardware manufacturing scalability: Producing humanoid robots at the volumes required to generate meaningful revenue requires supply chain relationships, manufacturing tooling, and quality control processes that take years to establish. The company's current funding provides runway, but manufacturing scale-up is a different challenge from research and demonstration.

Competition for the same market: BYD, CATL, and other major Chinese manufacturers are simultaneously investing in their own automation capabilities and evaluating multiple humanoid robot vendors. Tashi Zhihang is not the only company pursuing wire harness assembly automation in China.

Customers & deployments

This module is being compiled — no data to show yet.

08Markets and Use Cases

Where Tashi Zhihang Is Targeting Its Technology

Tashi Zhihang's stated commercial focus is narrow by design: smart manufacturing and intelligent logistics, with the wire harness assembly use case serving as both a technical proof-point and a market entry wedge ¹². This specificity is strategically coherent. Wire harness assembly is one of the most labour-intensive, injury-prone, and automation-resistant tasks in automotive and electronics manufacturing. It requires dexterous manipulation of flexible, deformable objects — precisely the class of problem that has defeated conventional industrial robots for decades and that the company's tactile-sensing and sub-millimetre precision claims are engineered to address.

Wire Harness Assembly: The Anchor Use Case

The global wire harness market is substantial and structurally dependent on manual labour. Automotive wire harnesses in particular have resisted full automation because the cables are limp, the connectors are small, and the assembly sequences are complex and variant-rich. A robot capable of reliably performing 105 sub-millimetre connector insertions per hour — the figure Tashi Zhihang demonstrated for its Guinness record in March 2026 ⁶ — would represent a meaningful productivity threshold if sustained across shifts and product variants. The company reports this capability has entered "industrial verification stage" ², which in Chinese manufacturing parlance typically means a controlled pilot at a partner facility, not volume production deployment.

The choice of wire harness assembly as a flagship use case also has a signalling function for investors and potential customers. It is a task that is independently understood to be hard, it has a measurable output metric (assemblies per hour), and it is amenable to a Guinness record format that generates media coverage. Whether the demonstration conditions — fixture design, cable preparation, lighting, environmental control — match those of a real production line is not publicly disclosed.

Smart Manufacturing: Broader Scope

Beyond wire harness assembly, the company positions its A-series and T-series humanoid robots for broader smart manufacturing applications ⁶⁷. The humanoid form factor is relevant here: factories designed for human workers do not require retooling to accommodate a robot with human-scale reach, grip, and mobility. This is the standard argument for humanoid robots in manufacturing, and it is commercially plausible in principle. However, the practical barriers — cycle time, reliability, mean time between failures, integration with existing MES and ERP systems, safety certification — are substantial and none of these have been publicly addressed by Tashi Zhihang in available sources.

Intelligent Logistics

Logistics is the second stated domain ¹². The specific applications are not detailed in available sources. EDITORIAL INFERENCE: The logistics use case likely encompasses intra-facility material handling, picking, and kitting tasks — the same territory contested by Boston Dynamics Spot and Stretch, Agility Robotics Digit, and a range of Chinese competitors including Unitree and Fourier Intelligence. Tashi Zhihang has not, to this analyst's knowledge, publicly demonstrated a logistics-specific capability distinct from its manufacturing demonstrations.

Geographic Market

All available evidence points to China as the primary and near-term exclusive market. The investor base is predominantly Chinese, the pilot partners are unnamed but implied to be domestic manufacturers, and the regulatory and procurement environment favours domestic suppliers ⁴⁹. Export ambitions, if any, are not publicly stated.

Addressable Market Sizing

UNKNOWN: Tashi Zhihang has not published a total addressable market estimate or a market sizing methodology in available sources. Third-party estimates for the global humanoid robot market vary enormously and are of limited analytical value at this stage of the industry's development.

---

09Competitive Landscape

Tashi Zhihang in the Field

Tashi Zhihang operates in one of the most intensely funded segments of the global technology industry in 2025-2026. The competitive field spans Chinese domestic rivals, US-based humanoid startups, and established industrial automation incumbents. The company's positioning — full-stack, software-and-hardware, manufacturing-focused — places it in direct competition with several well-capitalised peers.

Competitive comparison

Robot	Maker	Autonomy	Conf.
iRobot Roomba Combo 10 Max	iRobot	Autonomous	0.90
Mobile ALOHA (Stanford)	Stanford University	Teleoperated	0.90
1X NEO	1X Technologies	Remote-Assisted	0.90

Chinese Domestic Competition

The most directly comparable Chinese competitors are those pursuing humanoid robots for industrial applications with significant venture backing. Unitree Robotics (宇树科技) is the most prominent, having achieved commercial sales of its G1 and H1 humanoid platforms and built a substantial international profile ³⁰. Unitree's approach differs: it prioritises hardware cost reduction and broad platform availability over a single high-precision use case. Fourier Intelligence (傅利叶智能) targets rehabilitation and industrial applications with its GR-1 and GR-2 platforms and has disclosed customer deployments. Agility Robotics' Chinese-market positioning is limited, giving domestic players room.

Other Chinese embodied AI startups that have raised significant capital in the same period include AgiBot (智元机器人), Galbot (银河通用), and Astribot (星动纪元). Each has a distinct technical emphasis: AgiBot on dexterous manipulation, Galbot on mobile manipulation for logistics, Astribot on high-speed manipulation. The competitive dynamic among these companies is primarily for talent, pilot customers, and investor attention rather than for market share in a mature commercial market that does not yet exist at scale.

US and International Competition

Figure AI, Physical Intelligence (pi), Apptronik, and 1X Technologies represent the US competitive set. Physical Intelligence in particular has published research on generalised robot learning policies (pi0, pi0-FAST) that overlaps with Tashi Zhihang's AWE 3.0 ambitions. Boston Dynamics remains the benchmark for hardware reliability and has the advantage of Hyundai's manufacturing and customer relationships. None of these companies are currently competing directly in the Chinese domestic manufacturing market, but their research outputs set the global technical standard against which Tashi Zhihang's claims must be assessed.

Competitive Positioning Analysis

The "Full-Stack" Claim

Tashi Zhihang's insistence on full-stack ownership — developing core components, sensors, and AI models in-house — is a strategic choice with both advantages and risks. The advantage is integration: a company that controls its own actuators, sensors, and models can optimise across the stack in ways that a systems integrator cannot. The risk is capital intensity and the difficulty of achieving best-in-class performance at every layer simultaneously. EDITORIAL INFERENCE: At 14 months of age, Tashi Zhihang cannot plausibly have achieved genuine best-in-class performance across all hardware and software layers. The full-stack claim is better understood as a statement of strategic intent and a barrier to dependency on suppliers who might be subject to export controls or geopolitical disruption.

The Guinness Record as Competitive Positioning

The decision to pursue a Guinness World Record for wire harness assembly is a deliberate competitive move. It creates a specific, named, independently certified benchmark that competitors must respond to or surpass. It also frames the competitive conversation around precision manipulation rather than locomotion or general-purpose dexterity, where Tashi Zhihang would be at a disadvantage relative to more mature platforms. EDITORIAL INFERENCE: This framing is strategically intelligent but should not be mistaken for evidence of broad manufacturing readiness.

---

10Geopolitical Context and Constraints

Operating at the Intersection of Technology Competition and Industrial Policy

Tashi Zhihang's emergence cannot be understood without reference to the geopolitical environment in which it was founded and funded. The company was established in February 2025, a period of intensifying US-China technology competition, active Chinese state industrial policy in robotics and AI, and growing investor appetite for domestic alternatives to US-controlled technology stacks.

Chinese State Industrial Policy

The Chinese government has identified humanoid robots and embodied AI as strategic priorities under its "Made in China 2025" successor frameworks and the 14th Five-Year Plan for robotics. State-backed funds participated in Tashi Zhihang's Pre-A round: the Beijing Robot Industry Development Fund and Shanghai State Investment Pioneer are both listed as investors ¹¹²⁸. This is described in available sources as the first-ever embodied-AI investment by these state funds ²⁸, which signals both the novelty of the sector and the government's intention to support it. State participation in a funding round carries implications beyond capital: it provides a degree of political protection, facilitates access to state-owned enterprise customers, and aligns the company's interests with national technology objectives.

Export Controls and Supply Chain Risk

The US government's progressive tightening of export controls on advanced semiconductors, AI accelerators, and related components creates a structural constraint for any Chinese AI hardware company. Tashi Zhihang's emphasis on in-house component development and its "full-stack" positioning ⁶⁷ can be read partly as a response to this risk: a company that does not depend on NVIDIA H100s or other controlled components for its inference stack is less vulnerable to supply disruption. UNKNOWN: The specific semiconductor and component dependencies of Tashi Zhihang's hardware platform are not publicly disclosed. It is not possible to assess the company's actual exposure to export control risk from available sources.

Talent and Intellectual Property

The leadership team's backgrounds — Chen Yilun from Huawei, Li Zhenyu from Baidu, Ding Wenchao as Chief Scientist — reflect the concentration of AI and robotics talent in Chinese technology companies ⁹³⁰. Huawei's experience navigating US sanctions has produced a generation of engineers with specific expertise in designing around component restrictions. Whether this expertise translates to robotics hardware is an open question, but the institutional knowledge is relevant.

International Market Access

EDITORIAL INFERENCE: Tashi Zhihang's current focus on the Chinese domestic market is partly a function of its early stage and partly a function of the geopolitical environment. Chinese humanoid robots face potential regulatory scrutiny in Western markets on national security grounds — the same concerns that have affected Huawei telecommunications equipment and DJI drones. The company has not publicly articulated an international market strategy, and it would be premature to assume one is imminent.

The Meituan Connection

Meituan's participation as a lead investor in both the angel+ and Pre-A rounds ³⁴ is strategically significant beyond the capital it provides. Meituan operates one of China's largest logistics and delivery networks. Its investment in Tashi Zhihang suggests a potential future customer relationship for logistics robotics applications, and it provides the startup with access to real-world operational data and deployment environments that would otherwise be difficult to obtain. This is a form of strategic alignment that is common in Chinese technology investment but should be noted as a potential source of customer concentration risk if the relationship deepens.

Valuation and the Financing Environment

The speed and scale of Tashi Zhihang's fundraising — approximately $697 million in 14 months ²⁸ — reflects a specific moment in Chinese venture capital: a period of intense competition among top-tier funds for positions in embodied AI, driven by the belief that the sector is approaching an inflection point analogous to large language models in 2022-2023. Whether this belief is correct is unknowable at present. What is observable is that the financing environment has allowed Tashi Zhihang to accumulate capital at a rate that would have been impossible in a more sceptical market. The risk is that the capital creates expectations — of product milestones, customer deployments, and revenue — that the technology timeline may not support.

---

11The Hype, the Real and the Ugly

Separating Signal from Noise in Tashi Zhihang's Public Narrative

Tashi Zhihang has generated an exceptional volume of media coverage for a company that is pre-revenue, less than 18 months old, and has not disclosed a single named paying customer. This section applies the evidence discipline established in this report's preface to the company's principal public claims.

Claim tracker

Tashi Zhihang's A1 robot set a Guinness World Record for sub-millimeter wire harness assembly at 105 assemblies per hour in March 2026.Unknown

Multiple sources [6][7][9][10] report the Guinness record on March 10, 2026, but all trace back to company announcements or vendor-affiliated media; no independent Guinness adjudication report or third-party factory audit has been cited to confirm the record or the 105/hour figure in a sustained production context.

AWE 3.0 includes a built-in Failure Recovery mechanism via latent space future outcome prediction.Unknown

This capability is described consistently across vendor-affiliated sources [6][7][9] but no independent benchmark, peer-reviewed paper, or third-party test has been cited that validates the failure recovery mechanism's performance or reliability in real tasks.

Tashi Zhihang has accumulated 100,000+ hours of high-quality operational data via its SenseHub data collection suite.Unknown

Two dossier sources corroborate the 100,000+ hours figure, but both appear to be vendor-reported or vendor-affiliated; no independent audit, third-party data repository, or research publication has verified the volume or quality of this dataset.

Affiliated research papers demonstrate meaningful capability advances: RobMRAG (+7.76% over zero-shot baseline), TapSampling (policy-agnostic improvement without finetuning), Tac-Man (near-perfect tactile-only manipulation of articulated objects), and EmboAlign (+43.3pp over strongest baseline, zero-shot).Unknown

The four papers are available on arXiv [18][19][20][21] and represent independent academic artifacts, but they are affiliated research (not third-party evaluations of TARS products), the results are self-reported within the papers, and none has been confirmed as peer-reviewed/published in a venue that independently validated the benchmarks against TARS's deployed hardware.

Tashi Zhihang raised ~$697M across three rounds in ~14 months (angel ~$120M, angel+ $122M, Pre-A $455M), setting successive Chinese embodied-AI financing records, with a post-money valuation of ~$1.88B–$3B.Supported

The Pre-A $455M round is independently reported by TechInAsia [3], EqualOcean [4], MarketScreener [1][13], and multiple financial outlets; the angel rounds are corroborated by Caproasia [12][17] and Tracxn [29], though the valuation range ($1.88B vs $3B) remains conflicted between two independent market data sources and one commentary piece.

Claim 1: "World's First Native Embodied Foundation Model with Tri-Modal Alignment"

COMPANY CLAIM. The designation "world's first" for AWE 3.0 is unverified marketing language ⁶⁷. Physical Intelligence's pi0 model, Figure AI's Helix, and several academic research groups have developed multimodal robot foundation models with vision, language, and action components. The specific combination Tashi Zhihang describes — "native" tri-modal alignment of visual perception, language understanding, and motion control — may have a precise technical definition that distinguishes it from competitors, but no independent technical comparison has been published. The claim should be treated as vendor positioning, not established fact.

Claim 2: Guinness World Record for Sub-Millimetre Wire Harness Assembly

VERIFIED FACT (with caveats). The Guinness World Record for "most sub-millimetre wire harness assemblies by a robot in one hour" was set on March 10, 2026, at 105 assemblies/hour using the A1 robot with AWE 3.0 ⁶⁷. Guinness certification is an independent process that verifies the specific claim as stated. The caveats are important: Guinness records verify performance under the conditions of the attempt, not under production conditions. The fixture design, cable preparation, environmental control, and task definition are set by the record applicant. A record of 105 assemblies/hour in a controlled demonstration does not establish that the robot can achieve comparable throughput on a real production line with variant cables, worn connectors, and the interruptions of a factory environment.

Claim 3: "Industrial Verification Stage" for Wire Harness Assembly

COMPANY CLAIM, partially corroborated. The 36kr source ² reports that sub-millimetre wire harness assembly capability has entered industrial verification stage. This is consistent with the company's overall stage description (pre-revenue, pilot). "Industrial verification" in Chinese manufacturing typically means a controlled pilot at a partner facility, with defined success criteria, before volume deployment. It does not mean the technology is production-ready or that a commercial contract exists. No partner facility is named in available sources.

Claim 4: 100,000+ Hours of Operational Data

COMPANY CLAIM, reported by two sources ⁶⁹. The figure refers to data collected via the SenseHub suite (data collection gloves and first-person cameras). The quality, diversity, and relevance of this data to production tasks is not independently assessed. Data volume is a necessary but not sufficient condition for capable robot learning; the curation, labelling, and task coverage of the dataset matter as much as raw hours. UNKNOWN: The methodology for counting "high-quality operational data hours" is not disclosed.

Claim 5: WIYH Dataset as "World's First Large-Scale Real-World VLTA Dataset"

COMPANY CLAIM. The WIYH (World In Your Hands) dataset is described as the world's first large-scale real-world embodied Vision-Language-Tactile-Action multimodal dataset, and is reported to have been open-sourced ⁶. The "world's first" designation for the specific VLTA combination is plausible but unverified. The dataset's open-source status, if confirmed, would be a genuine contribution to the research community and would allow independent assessment of its quality and coverage. UNKNOWN: The repository location, licence terms, and size of the WIYH dataset are not specified in available sources.

Claim 6: Post-Money Valuation of ~$3 Billion

CONFLICTED. Two independent market data sources (MarketScreener UK and India editions) cite a post-money valuation of CNY 13 billion (~$1.88 billion USD) ¹¹³. A single commentary source cites ~$3 billion (CNY 20 billion) ²⁸. The lower figure is better supported by independent sources. The discrepancy may reflect different valuation methodologies, different timing, or the inclusion of secondary transactions. This report uses the CNY 13 billion figure as the more defensible estimate while noting the conflict.

The Ugly: What Is Not Being Said

Several important facts are absent from Tashi Zhihang's public narrative and deserve explicit attention:

• No named customers. In 14 months of operation and across three funding rounds, no paying customer has been publicly identified. The "industrial verification" partner is unnamed.
• No reliability data. The Guinness record establishes a peak performance figure. Mean time between failures, error rates across cable variants, and performance degradation over time are not disclosed.
• No independent technical review. The AWE 3.0 model has not been subjected to independent benchmarking. The research papers ^18192021 are affiliated work, not independent audits.
• No revenue timeline. The company has not publicly committed to a revenue milestone or a commercial launch date.
• Founder track records are relevant but not sufficient. Chen Yilun's experience at Huawei and Li Zhenyu's at Baidu are genuine credentials ⁹³⁰, but neither has previously built a robotics hardware company from scratch. The operational challenges of manufacturing humanoid robots at scale are distinct from those of software or telecommunications product development.

---

12Future Scenarios

Three Plausible Trajectories for Tashi Zhihang

Given the evidence available as of June 2026, three scenarios bracket the company's likely development over the next 24-36 months. These are not predictions; they are structured assessments of outcomes consistent with the available evidence.

Scenario A: Controlled Execution — The Narrow Champion

Probability: Moderate

In this scenario, Tashi Zhihang successfully converts its wire harness assembly capability into a commercial product deployed at two to four named manufacturing customers in China by end of 2027. Revenue is modest — single-digit millions of USD — but the deployments provide the operational data and reliability evidence needed to expand the use case portfolio. The AWE 3.0 model improves iteratively through production deployment. The company raises a Series A at a valuation consistent with the CNY 13 billion post-Pre-A figure, with participation from one or more strategic industrial investors. International expansion is deferred.

This scenario requires: successful completion of the current industrial verification pilot; a commercial contract with at least one named customer; demonstrated reliability over multi-shift operation; and continued access to capital at reasonable dilution. None of these are guaranteed, but none are implausible given the company's resources and the strength of its investor syndicate.

Scenario B: Platform Expansion — The Embodied AI Contender

Probability: Lower, but not negligible

In this scenario, the wire harness assembly use case proves to be a genuine beachhead. The AWE 3.0 model generalises more effectively than current evidence suggests, and the company deploys across multiple manufacturing task categories — assembly, inspection, material handling — within 24 months. Meituan's logistics network becomes an early customer for intra-facility robot deployment. The company achieves eight-figure annual revenue by end of 2027 and positions for a Series B or pre-IPO round. The WIYH dataset and research publications attract top-tier academic collaborations that accelerate capability development.

This scenario requires: genuine generalisation capability in AWE 3.0 beyond the demonstrated wire harness task; operational reliability at production scale; and a commercial relationship with Meituan or a comparable anchor customer. The research papers ^18192021 provide some evidence that the technical foundations for generalisation exist, but the gap between research results and production deployment is large.

Scenario C: Capital Consumption Without Commercialisation

Probability: Non-trivial

In this scenario, the technical challenges of production-grade humanoid manipulation prove more intractable than the funding pace implies. The industrial verification pilot extends without conversion to a commercial contract. The $697 million in raised capital provides a long runway — likely three to five years at plausible burn rates — but the company fails to achieve the revenue milestones that would justify a Series A at current valuations. The broader Chinese embodied AI funding environment cools as investors reassess timelines. The company either pivots to a narrower software licensing model, merges with a larger industrial automation player, or enters a prolonged pre-revenue holding pattern.

This scenario is consistent with the historical pattern of robotics startups that have raised large early rounds on the strength of impressive demonstrations but struggled to bridge the gap to production reliability. It does not require any specific failure; it requires only that the technology timeline be longer than the investor narrative assumes.

Key Inflection Points to Watch

The transition between these scenarios will be signalled by a small number of observable events:

1. Announcement of a named commercial customer with a disclosed contract value or deployment scale.
2. Publication of independent technical benchmarking of AWE 3.0 against peer models.
3. Disclosure of the industrial verification pilot partner and its assessment of the technology.
4. Evidence of multi-shift, multi-variant production operation (not a demonstration).
5. A Series A round at a valuation that confirms or revises the current CNY 13 billion figure.

---

13What to Watch: A Live Monitoring Checklist

The following indicators, organised by category, constitute the minimum evidence set that would materially change this report's assessments. Analysts and investors tracking Tashi Zhihang should update their views when these signals appear.

Commercial Milestones

• First named paying customer publicly disclosed, with deployment scope described
• Commercial contract value or robot unit count disclosed (even in a range)
• Industrial verification pilot partner identified and pilot outcome reported
• Revenue figure disclosed in any regulatory or investor document
• Second named customer, indicating repeatability rather than a single bespoke deployment

Technical Milestones

• AWE 3.0 benchmarked independently against Physical Intelligence pi0, Figure Helix, or equivalent
• WIYH dataset repository made publicly accessible with documented size, task coverage, and licence
• A1 or successor robot demonstrated on a second distinct manipulation task at comparable precision
• Mean time between failures or error rate data disclosed for any production or pilot deployment
• Peer-reviewed publication (not arxiv preprint) from Tashi Zhihang-affiliated authors in a top-tier venue (ICRA, CoRL, NeurIPS, ICLR)

Funding and Corporate

• Series A round announced, with lead investor and valuation disclosed
• Valuation conflict resolved: CNY 13 billion vs CNY 20 billion figure clarified by a primary source
• State-owned enterprise customer announced (would confirm state-fund investment thesis)
• Meituan commercial deployment announced (would confirm strategic investor thesis)
• Any regulatory filing, patent grant, or product certification in China or internationally

Competitive Signals

• A Chinese competitor achieves a comparable or superior wire harness assembly benchmark, challenging the Guinness record's competitive significance
• A US competitor (Physical Intelligence, Figure, Apptronik) announces a China market entry or partnership
• Tashi Zhihang poaches or loses key technical staff (signals capability trajectory)
• Any independent teardown or technical review of A1 or T-series hardware

Risk Signals

• Industrial verification pilot reported as unsuccessful or extended without explanation
• Key founder departure (Chen Yilun, Li Zhenyu, or Ding Wenchao)
• Export control action affecting components used in A1 or T-series hardware
• Funding round delayed or downsized relative to market expectations
• Media reports of safety incidents during pilot operations

---

14Sources and Methodology

Source List

¹ Shanghai Tashi Zhihang Technology Co., Ltd. announced that it has received $455 million in funding from a group of investors | MarketScreener UK — https://uk.marketscreener.com/news/shanghai-tashi-zhihang-technology-co-ltd-announced-that-it-has-received-455-million-in-funding-f-ce7e50dddd8af320

² HSG and Hillhouse Capital Team Up Rarely: $455M Bet on a "Brain" — https://eu.36kr.com/en/p/3778900472780037

³ Chinese embodied AI startup Tars lands $122m funding — https://www.techinasia.com/news/chinese-embodied-ai-startup-tars-lands-122m-funding

⁴ Chinese embodied AI Company TARS raised $455 million in a Pre-A round, breaking the record for the sector in China, led by Hillhouse, Sequoia, and Meituan | EqualOcean — https://equalocean.com/news/2026041621834-chinese-embodied-ai-company-tars-raised-455-million-pre-round-breaking-record

⁵ AI costs spike as subscriptions hit pricing wall — Tom's Hardware — https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-costs-spike-as-subscriptions-hit-pricing-wall-firms-turn-towards-chinese-llms-open-source-models-to-extend-budget

⁶ Tashi Zhihang AWE 3.0 Raises $455M, Setting New Embodied AI Funding Record | Embodied Global — https://embodiedglobal.com/en/article/tashi-zhineng-awe3-455m-funding-may-2026

⁷ TARS AI Closes $455M Pre-A Round, Setting China Embodied-AI Financing Record — RobotToday — https://robottoday.com/article/tars-ai-closes-455-m-pre-a-round-setting-china-embodied-ai-financing-record

⁸ Shanghai Tashi Zhihang Technology Co., Ltd. announced that it has received $455 million in funding from a group of investors | MarketScreener Hong Kong — https://hk.marketscreener.com/news/shanghai-tashi-zhihang-technology-co-ltd-announced-that-it-has-received-455-million-in-funding-f-ce7e50dddd8af320

⁹ Tars Raises $455M: How China's 'Robot Brain' Startup Cracked the Embodied Intelligence Code | AI in China — https://www.ainchina.com/blog/tars-embodied-intelligence-455-million-brain-club/

¹⁰ China's TARS AI Raises $455M in Embodied Intelligence Round — https://olachina.org/tars-ai/

¹¹ "Tashi Intelligent Navigation" Raises $455 Million in Pre-A Round, Linear Capital Continues to Back Through Three Consecutive Rounds Since Seed Stage | Linear Portfolio | elsewhere — https://elsewhere.news/en/linearcapital/455pre-a-linear-portfolio

¹² China AI Tech Startup TARS (Tashi Zhihang) Raised $120 Million in Fund Raising Round, Founded in 2025 February | Caproasia — https://www.caproasia.com/2025/03/28/china-ai-tech-startup-tars-tashi-zhihang-raised-120-million-in-fund-raising-round-founded-in-2025-february-investors-include-lanchi-ventures-qiming-venture-partners-linear-capital-hengxu-capit/

¹³ Shanghai Tashi Zhihang Technology Co., Ltd. announced that it has received $455 million in funding from a group of investors | MarketScreener India — https://in.marketscreener.com/news/shanghai-tashi-zhihang-technology-co-ltd-announced-that-it-has-received-455-million-in-funding-f-ce7e50dddd8af320

¹⁴ Shoucheng Holdings fund invests in Chinese embodied intelligence firm Tashi Zhihang — https://app.dealroom.co/news/feed/shoucheng-holdings-fund-invests-in-chinese-embodied-intelligence-firm-tashi-zhihang

¹⁵ AI Market Watch's Post — LinkedIn — https://www.linkedin.com/posts/ai-market-watch_tashi-zhihang-%E5%AE%83%E7%9F%B3%E6%99%BA%E8%88%AA-raises-455m-pre-a-to-activity-7460832152792813569-YnRm

¹⁶ Embodied intelligence startup Tashi Zhihang completes $120 million funding — https://en.eeworld.com.cn/mp/AIxintianxia/a396116.jspx

¹⁷ China AI Tech Startup TARS (Tashi Zhihang) Raised $120 Million in Fund Raising Round — https://www.caproasia.com/2025/03/28/china-ai-tech-startup-tars-tashi-zhihang-raised-120-million-in-fund-raising-round-founded-in-2025-february-investors-include-lanchi-ventures-qiming-venture-partners-linear-capital-hengxu-capit

¹⁸ Zero-Shot Robotic Manipulation via 3D Gaussian Splatting-Enhanced Multimodal Retrieval-Augmented Generation — https://arxiv.org/html/2603.00500v1

¹⁹ TapSampling: Inference-Time Sampling with a Task-Progress-Understanding Verifier for Robotic Manipulation — https://arxiv.org/html/2605.25547

²⁰ Tac-Man: Tactile-Informed Prior-Free Manipulation of Articulated Objects — https://arxiv.org/html/2403.01694v1

²¹ EmboAlign: Aligning Video Generation with Compositional Constraints for Zero-Shot Manipulation — https://arxiv.org/abs/2603.05757v1

²² They cant keep this private — Samsung S26 Ultra Teardown — https://www.youtube.com/watch?v=TRW4W7KkJXs [NOT APPLICABLE — misattributed source]

²³ MacBook Neo Teardown: The Most Repairable Mac Yet? — https://www.youtube.com/watch?v=5k7Lv7f-5CQ [NOT APPLICABLE — misattributed source]

²⁴ Is This The Perfect Laptop/Tablet Hybrid? #framework #ifixit #teardown — https://www.youtube.com/watch?v=wY-z6jqgBok [NOT APPLICABLE — misattributed source]

²⁵ Shop Vac Micro 1 Gallon Review & Teardown — https://www.youtube.com/watch?v=uaL44q17rQU [NOT APPLICABLE — misattributed source]

²⁶ Pushing the Fosi Audio SW10 to the Limit: Driver Excursion Demo — https://www.youtube.com/watch?v=ElsjB0ICMFI [NOT APPLICABLE — misattributed source]

²⁷ This Android Phone is ALSO a Linux AND Windows PC! NexPhone Hands On! — https://www.youtube.com/watch?v=qfY5t6N8YxM [NOT APPLICABLE — misattributed source]

²⁸ Over 3 billion! The birth of China's largest embodied intelligent financing to date — Taibo — https://en.taibo.cn/p/26111433

²⁹ Tashi Zhihang — 2026 Company Profile, Funding & Competitors — Tracxn — https://tracxn.com/d/companies/tashizhihang/__eEYP1f1rvZQFGX1ZsH7SzG9Aj69ajF-n8lPUHseIies

³⁰ TARS Robotics (Trusted AI and Robotics Solution or Tashi Zhihang in Chinese), Shanghai, China | Portal of Robotics and Artificial Intelligence — https://pr.ai/threads/tars-robotics-trusted-ai-and-robotics-solution-or-tashi-zhihang-in-chinese-shanghai-china.27346/

Methodology Note

This report was produced under the Max Robotics Premium Editorial standard. Evidence is classified into four categories throughout the text:

| Label |