Chapter 13: China's Leaders: Unitree and AgiBot

13.1 Why China's humanoid leaders matter

Chapter 11 (Boston Dynamics) and Chapter 12 (Figure · Agility · Tesla) catalogued the US trio's architectural bets. Chapter 13 turns to China's two dominant humanoid companies — Unitree Robotics (founded 2016, Hangzhou) and AgiBot (founded 2023, Shanghai) — whose architectural choices, commercial models, and deployment trajectories are distinct from the US pattern in ways that matter to the 2026–2028 trajectory.

Three structural differences frame the chapter. First, pricing: Unitree's G1 ships at approximately $16,000 USD base [Unitree, 2024-g1], which is one to two orders of magnitude below US-market humanoid prices ($30k–$200k range depending on specification). AgiBot's A2 is in the mid-$30k range. Second, hardware openness: Unitree's G1 ships with open-source unitree_rl_gym training frameworks and published joint specifications, meaning third-party researchers can reproduce Unitree's baseline policies — a posture US vendors do not share. Third, data-platform scale: AgiBot World Colosseo ^[5] is the largest open manipulation dataset released by a humanoid company (1M+ trajectories across 217 tasks on 100 physical humanoid units), rivalling Open X-Embodiment in scope but concentrated on a single embodiment family.

The chapter proceeds: Unitree's hardware lineage and G1/H1 platforms (§13.2), the Unitree open-RL-framework posture and academic adoption (§13.3), AgiBot's emergence and A1/A2 platforms (§13.4), the AgiBot World Colosseo data infrastructure and its Genie Sim 3.0 partner simulator (§13.5), AgiBot's GO-1 and GO-2 VLA architectures (§13.6), Fourier Intelligence GR-1/GR-2 and Xiaomi CyberOne as secondary Chinese players (§13.7), a Unitree-vs-AgiBot architectural contrast (§13.8), and open questions about how the China ecosystem will evolve relative to US and Korean counterparts (§13.9).

13.2 Unitree Robotics — hardware lineage and platforms

Unitree Robotics was founded in 2016 by Xingxing Wang, a former Alibaba and DJI engineer. The company built its initial reputation on quadrupeds — Laikago (2017), Aliengo (2019), A1 (2020), Go1 (2021), Go2 (2023) — positioning itself as the Chinese counterpart to Boston Dynamics's Spot but at radically lower price points (A1 at approximately $10,000 USD vs Spot's $74,500). The quadruped lineage gave Unitree manufacturing fluency and supplier relationships that it leveraged into humanoid platforms starting 2023.

Unitree H1 (August 2023) [Unitree, 2024-h1] was Unitree's first humanoid. Specifications:

• Height: 180 cm; mass: approximately 47 kg.
• Total DoF: 19 (6 per arm, 6 per leg, 1 torso).
• Walking speed: 1.5 m/s demonstrated, 3.3 m/s peak reported in controlled demos.
• Actuators: QDD-class quasi-direct-drive (Chapter 4 §4.4), Unitree-manufactured (not licensed from a third-party actuator vendor).
• Base price: approximately $90,000 USD at launch (2023) — substantially below US frontier-humanoid prices.

H1 was positioned by NVIDIA as one of the partner humanoid platforms alongside Fourier GR-1 and 1X Neo in GR00T N1 public announcements, though GR00T N1's published real-world evaluations were carried out on the Fourier GR-1 specifically ^[7]. Academic RL papers through 2024–2025 use H1 extensively: Radosavovic et al. ^[11] trains transformer policies on H1; HOVER ^[12] demonstrates versatile whole-body control on H1; FALCON ^[13] shows force-adaptive loco-manipulation on H1. The H1's combination of QDD hardware, mid-tier price, and Unitree's open training stack made it the default research humanoid for Chinese and international labs alike through 2024–2025.

Unitree G1 (June 2024) [Unitree, 2024-g1] was Unitree's follow-up and the platform that crystallized the low-price-point strategy. Specifications:

• Height: 127 cm (adjustable to 132 cm); mass: approximately 35 kg.
• Total DoF: 23 (base), 43 (with 3-finger hands).
• Actuators: QDD with wider torque range than H1; joint specifications published.
• Base price: approximately $16,000 USD — roughly one-fifth H1's launch price and less than one-tenth Figure 03 or Electric Atlas equivalents.

G1's $16k price point changed the humanoid research-lab calculus. Labs that could not justify a $100k+ research platform could justify a G1 plus a researcher's salary at equivalent cost. The platform's compact size (the 127 cm form factor is roughly child-proportioned, not adult-proportioned) is a deliberate trade-off: it reduces mass and actuator cost, at the cost of payload and end-effector reach. For academic research focused on locomotion RL, whole-body coordination, and sim-to-real methodology, the trade-off favored G1 adoption.

Unitree's 2025–2026 follow-ups include H1-2 (heavier payload, revised actuators) and the announced 2026 H2 platform, pushing toward heavier-payload industrial use cases while maintaining pricing pressure on the broader market.

13.3 Unitree's open-RL-framework posture

What distinguishes Unitree structurally from US frontier humanoid companies is the degree of hardware-and-software transparency. Three dimensions:

Published joint specifications: Unitree publishes torque limits, gear ratios, encoder resolutions, and URDF models for G1 and H1. US vendors (Figure, Tesla, 1X) treat joint specifications as competitive IP. The Unitree posture supports third-party sim-to-real research directly — a researcher with Isaac Lab and Unitree's URDF can reproduce a policy before buying the hardware.

Open-source unitree_rl_gym: Unitree maintains unitree_rl_gym, an Isaac Gym / legged_gym-based training stack adapted for Unitree humanoids [Unitree, 2024-g1]. The repository ships with baseline policies trained on flat ground, stairs, and outdoor-terrain curricula. Humanoid-Gym ^[14] and Booster Gym ^[15] build on the same lineage. The sim-to-real pipeline documented in §5.3 — Isaac Gym → MuJoCo → hardware — has Unitree as the exemplar.

Public commercial pricing and ordering: Unitree's G1 and H1 are purchasable from the company website with published prices and lead times. The US frontier humanoid market is quote-based and relationship-gated; Unitree's market model is closer to consumer-electronics ordering.

The strategic implication is that Unitree is seeding the global humanoid research base. Papers published at CoRL, IROS, ICRA, RSS, and major workshops in 2024–2025 increasingly run experiments on Unitree hardware (H1, G1, or Go2) because it is the most cost-accessible research platform. Whether this pipeline advantage converts into long-term commercial dominance depends on AgiBot, Figure, and others' ability to match Unitree's pricing at comparable capability levels — a test that the 2026–2028 window will run.

13.4 AgiBot — the 2023 entrant and the A1/A2 lineage

AgiBot (智元機器人) was founded in February 2023 by Peng Zhihui (formerly of Huawei's Genius Youth program) in Shanghai. Within two years, AgiBot emerged as the most data-platform-focused humanoid company in China, distinguished from Unitree's hardware-first positioning by its focus on fleet-scale data collection and VLA model releases.

AgiBot A1 (late 2023 / early 2024) was AgiBot's first humanoid, a 175 cm, ~53 kg platform with 41 DoF. A1 was less a product than a research scaffold for AgiBot World Colosseo data collection (§13.5).

AgiBot A2 (2025) is the commercial-grade humanoid. Specifications from public disclosures:

• Height: approximately 175 cm; mass: approximately 55 kg.
• Total DoF: 49 (including 12-DoF two-finger-plus hands).
• Walking speed: reportedly 1–2 m/s continuous.
• Price: approximately $35,000 USD — positioned between Unitree H1 and US frontier-price range.

AgiBot A2 targets industrial deployment, consumer services, and data-collection teleoperation in roughly equal measure. The platform is a less-refined industrial-only robot than Digit but more manipulation-capable than Unitree G1. The architectural bet is that A2 is the data-collection endpoint — AgiBot's real product is the fleet-learning pipeline documented in §13.5 and §13.6, with A2 being the physical substrate.

13.5 AgiBot World Colosseo and Genie Sim 3.0

AgiBot's strategic distinctive is data infrastructure, and the 2025 release of AgiBot World Colosseo ^[5] is the largest open humanoid manipulation dataset released by any humanoid company through 2026Q1.

Scale:

• 100 physical humanoid units performing real-world manipulation simultaneously across AgiBot's Shanghai data-collection facility.
• 1,000,000+ trajectories collected and released.
• 217 distinct tasks spanning kitchen, office, retail, warehouse, and household domains.
• Multi-modal recordings: robot state, joint commands, two-camera RGB, wrist-camera RGB, depth, force/torque.

Comparison context: Open X-Embodiment ^[10] has 970,000 episodes across 22 embodiments — broader but shallower per-embodiment. AgiBot World is deeper on the A1/A2 embodiment specifically and therefore more useful for training single-embodiment specialized models. The dataset is released under a non-commercial research license.

Genie Sim 3.0 ^[6] is AgiBot's companion high-fidelity simulation platform, released in 2026. Genie Sim is specifically optimized for the A1/A2 embodiment with:

• Photorealistic rendering with physically-based materials.
• GPU-parallel rigid-body and soft-body physics (built on NVIDIA's PhysX / Isaac Sim lineage but with AgiBot's optimizations for humanoid contact).
• Pre-built scene libraries matching the 217 AgiBot World tasks.
• Sim-to-real bridge via Isaac Sim / MuJoCo hand-off (Chapter 5 §5.3 pattern).

The Colosseo-plus-Genie-Sim pairing is a vertically integrated data platform: real-world data collection on 100 robots feeds simulation, which feeds improved VLA training, which feeds back into the physical fleet. This loop is what AgiBot bets on as the durable competitive moat — a bet distinct from Figure's Helix end-to-end thesis, Agility's Motor Cortex industrial-narrow thesis, or Tesla's manufacturing-scale thesis.

13.6 AgiBot GO-1 and GO-2 — ViLLA and asynchronous dual-system

AgiBot has released two generations of VLA models to pair with the A2 hardware and AgiBot World data:

GO-1 (March 2025) ^[5] is positioned as "ViLLA" (Vision-Language-Latent-Action), a generalist embodied foundation model. The model is trained on AgiBot World's 1M+ trajectories, plus supplementary internet video and language data. Architecture details are less publicly disclosed than GR00T N1 or π0, but AgiBot's positioning emphasizes:

• A tokenized action representation mapping to A2's 49-DoF command space.
• Task conditioning via language instructions plus vision.
• Zero-shot transfer to held-out A2 tasks at reported >70% success on a benchmark suite.

GO-2 (2026) ^[6] advances to an asynchronous dual-system architecture, where:

• System 1 runs at 1 kHz (joint-level control, tightly coupled to hardware).
• System 2 runs at low frequency (semantic task-level planning).
• The two systems communicate asynchronously — event-driven messages rather than clock-synchronous handoffs. When S2 emits a revised task plan, S1 adopts it at the next convenient control tick; S1 does not block waiting for S2 decisions.

The asynchronous dual-system framing is architecturally distinct from Figure Helix's System 0/1/2 (synchronous within a learned weight set) and GR00T N1's hierarchical VLA-plus-diffusion-head. Chapter 9 §9.5 mapped this asynchrony against Figure's and GR00T's patterns; GO-2 is the explicit instantiation on production hardware. The quality trade-off is that S1 can continue operating even when S2 is compute-bound or network-delayed, at the cost that S1 operates on stale task plans during those intervals — acceptable for many industrial tasks but potentially dangerous for tasks where live task updating is safety-critical.

As of 2026Q1, GO-2's quantitative benchmarks on AgiBot World task success rates have not been released in peer-reviewed form; AgiBot's public claims emphasize deployment-scale evidence (multiple manufacturing pilots in Shanghai and Hangzhou) rather than per-task success numbers. The 2026–2027 release of GO-2 benchmarks will be a significant data point.

13.7 Secondary Chinese players — Fourier and Xiaomi

Beyond Unitree and AgiBot, three additional Chinese humanoid programs deserve mention:

Fourier Intelligence GR-1 / GR-2 [Fourier, 2024-gr1]: Founded in 2015 in Shanghai with initial rehabilitation-exoskeleton focus, Fourier pivoted to general-purpose humanoids with GR-1 (2023, 54 DoF, 165 cm, 55 kg) and GR-2 (2024, improved hands and torso). GR-1 is the NVIDIA GR00T N1 real-world deployment platform and is named alongside Unitree H1 and 1X Neo as NVIDIA partner humanoids in public GR00T N1 announcements. Fourier's architectural bet is research-platform breadth — selling GR-1/GR-2 to university labs globally at mid-tier pricing (~$90k–$120k USD range) and letting the research community develop domain-specific applications.

Xiaomi CyberOne ^[9]: Xiaomi announced CyberOne (2022) as a research humanoid at 177 cm, 52 kg. CyberOne was less commercially visible than Unitree or AgiBot products through 2024–2026 but signals Xiaomi's intent to enter the humanoid supply chain — consistent with Xiaomi's broader robotics strategy following its EV platform launch. Xiaomi's manufacturing capacity (comparable to Tesla at global scale) makes its humanoid trajectory worth tracking even though its current technical disclosure is thin.

Tencent and other entrants: Tencent RoboticsX has published research on bipedal locomotion and quadrupeds without commercial humanoid announcements through 2026Q1; XPeng Robotics (2024 entrant) announced humanoid intent but has thin public disclosure. These are "watching brief" entities as of the chapter's writing date.

The Chinese humanoid ecosystem, taken as a whole, is characterized by: (1) lower price points than US counterparts by roughly an order of magnitude; (2) stronger vertical integration with Chinese manufacturing supply chains; (3) less peer-reviewed academic disclosure relative to fielded unit count; and (4) faster iteration cycles driven by both regulatory and manufacturing advantages. The trajectory implication is that by 2028, China may have the largest deployed humanoid unit count even if US or Korean architectural leadership persists on specific capability axes (§13.8).

13.8 Unitree vs AgiBot — architectural contrast

A two-column comparison frames the intra-China split:

Three structural observations:

Hardware-first vs data-first. Unitree's competitive moat is low-cost, well-engineered, openly-documented humanoid hardware. AgiBot's moat is 100-robot fleet-scale manipulation data collection. These are different bets on what durably-accrues-value looks like in the humanoid industry. Both can succeed simultaneously — Unitree selling hardware to research labs and some industrial customers, AgiBot running a proprietary data-and-VLA business — but the long-term winner depends on whether open hardware or proprietary data compounds more strongly.

VLA vs control. Unitree does not release a VLA. The company's public stack is control-focused (sim-to-real RL for locomotion and whole-body coordination) and its VLA strategy, if any, is to partner with third parties (NVIDIA GR00T N1 on Unitree H1 is the exemplar). AgiBot has released two VLA generations (GO-1, GO-2) tightly coupled to its own hardware. The bet is different: Unitree lets the VLA-research ecosystem develop around its hardware, while AgiBot vertically integrates hardware-plus-VLA-plus-data.

Research adoption vs commercial deployment. Unitree's adoption metric is academic papers per year using Unitree hardware, which has grown rapidly 2023–2026 (hundreds of papers). AgiBot's adoption metric is manufacturing pilots deploying A2 plus GO-1/GO-2 downloads, which has a smaller absolute count but higher per-deployment integration depth. Both are legitimate metrics; neither dominates the other in 2026Q1.

13.9 Open questions

Three questions close Chapter 13:

First, will Unitree's open-hardware posture compound into a "Linux moment" for humanoid research? If the global research community converges on Unitree as the de facto platform — as happened with NVIDIA GPUs in deep learning and Linux in server operating systems — Unitree captures a durable ecosystem position that is difficult to displace even if other companies produce superior hardware. The alternative is that Unitree's openness is merely a cost-efficient acquisition strategy and that, once its installed base is large enough, the company restricts the openness to protect margin.

Second, is AgiBot World Colosseo a durable data moat or a temporary head start? Open X-Embodiment ^[10] is broader but shallower; 1X's home-robot consumer beta generates different data; AgiBot's 100-robot industrial fleet is specific and substantial but could be matched by a Figure fleet, Digit fleet, or Optimus fleet by 2028 if those companies execute on deployment. The moat is time-to-scale, not fundamental scarcity.

Third, how will US-China decoupling affect the humanoid technology stack? Unitree's export to US research labs has been smooth through 2026Q1 but is subject to policy risk. A Unitree that cannot sell into the US research market, or US research labs that cannot purchase Unitree hardware, would fragment the humanoid research ecosystem along geopolitical lines in ways that slow global progress. Chapter 14 and Chapter 16 return to geopolitical constraints as explicit variables in the 2026–2028 scenario analysis.

Chapter 14 pivots from the US-China axis to examine Korea's current humanoid position — the ecosystem, the national initiatives, the manufacturing supply chains that are differently-positioned from both US and Chinese peers.

References

1. Unitree Robotics. (2024). Unitree H1 humanoid platform specifications. Unitree product page and public technical disclosures, 2023–2024.
2. Unitree Robotics. (2024). Unitree G1 humanoid platform and `unitree_rl_gym` open-source training framework. Unitree product launch, June 2024.
3. AgiBot. (2025). AgiBot World Colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. AgiBot technical announcement, 2025.
4. AgiBot. (2026). Genie Sim 3.0: A high-fidelity comprehensive simulation platform for humanoid robotic learning. AgiBot technical announcement, 2026.
5. AgiBot. (2025). AgiBot GO-1: ViLLA generalist embodied foundation model. AgiBot technical announcement, March 2025.
6. AgiBot. (2026). AgiBot GO-2: Asynchronous dual-system humanoid control architecture. AgiBot technical announcement, 2026.
7. Bjorck, J., et al. (2025). GR00T N1: An open foundation model for generalist humanoid robots. NVIDIA technical report and arXiv preprint 2503.14734.
8. Fourier Intelligence. (2024). Fourier GR-1 / GR-2 humanoid platform. Fourier product announcements, 2023–2024.
9. Xiaomi. (2023). Xiaomi CyberOne humanoid platform. Xiaomi announcement, August 2022, subsequent updates through 2023.
10. Padalkar, A., et al. (2023). Open X-Embodiment: Robotic learning datasets and RT-X models. arXiv preprint 2310.08864.
11. Radosavovic, I., et al. (2024). Humanoid locomotion as next token prediction. NeurIPS. arXiv:2402.19469.
12. He, T., et al. (2024). HOVER: Versatile neural whole-body controller for humanoid robots. Proc. IEEE ICRA 2025. arXiv:2410.21229.
13. Zhang, Y., et al. (2025). FALCON: Learning force-adaptive humanoid loco-manipulation.
14. Gu, X., Zhang, Y., Wu, K., et al. (2024). Humanoid-Gym: Reinforcement learning for humanoid robot with zero-shot sim2real transfer. arXiv preprint 2404.05695.
15. Wang, Y., et al. (2025). Booster Gym: An end-to-end reinforcement learning framework for humanoid robot locomotion. arXiv preprint 2506.15132.