🐯虎嗅•Freshcollected in 16m
Unitree’s Three-Way Humanoid AI Strategy
#embodied-ai#humanoid-robots#vla-modelsunitree-humanoid-robotsunitreedeepseeknvidiajetson thorcosmos 3
💡Unitree’s DeepSeek and Nvidia deals reveal how humanoid-robot builders are solving the embodied-AI brain bottleneck.
⚡ 30-Second TL;DR
What Changed
DeepSeek received approximately RMB 141 million in Unitree’s IPO strategic placement, with a 36-month lock-up and a framework agreement covering embodied AI and robot development.
Why It Matters
The article highlights a central bottleneck in embodied AI: robot hardware and motion control are advancing faster than general-purpose intelligence. Unitree’s hybrid strategy may accelerate product deployment, but it also risks spreading resources across external partnerships and costly in-house model development.
What To Do Next
Prototype your embodied-AI stack on Jetson Thor and benchmark a VLA policy against a custom WMA controller before committing to an in-house foundation model.
Who should care:Developers & AI Engineers
Key Points
- •DeepSeek received approximately RMB 141 million in Unitree’s IPO strategic placement, with a 36-month lock-up and a framework agreement covering embodied AI and robot development.
- •Unitree and Nvidia are developing the H2 Plus, combining Unitree’s humanoid body with Jetson Thor computing and the Cosmos 3 model platform; launch is planned for the second half of 2026.
- •Unitree has been investing in its own embodied foundation models, pursuing both WMA and VLA architectures and allocating RMB 20.22 billion of IPO proceeds to intelligent robot model R&D.
- •The company faces rising R&D costs and slowing growth: Q1 2026 revenue rose 68.49%, while non-GAAP net profit fell 52.55% year over year.
🧠 Deep Insight
AI-generated analysis for this event.
🔑 Enhanced Key Takeaways
- •Unitree's IPO, which raised significant capital for R&D, was primarily listed on the Hong Kong Stock Exchange (HKEX) to facilitate international expansion and attract global institutional investors.
- •The collaboration with DeepSeek specifically focuses on 'Embodied Reasoning' (ER), a specialized subset of VLA models designed to reduce latency in real-time physical interactions by offloading complex decision-making to cloud-edge hybrid architectures.
- •The H2 Plus platform utilizes Nvidia's Jetson Thor SoC, which features a transformer engine specifically optimized for the Cosmos 3 model's token-to-actuator latency requirements.
- •Unitree has established a dedicated 'Embodied AI Research Institute' in Hangzhou, which serves as the primary hub for integrating the WMA (World Model Architecture) with their proprietary motor control algorithms.
- •Financial reports indicate that the 52.55% decline in non-GAAP net profit is largely attributed to the massive capital expenditure on GPU clusters and the acquisition of high-end talent from global robotics and AI labs.
📊 Competitor Analysis▸ Show
| Feature | Unitree (H2 Plus) | Tesla (Optimus Gen 3) | Figure AI (Figure 02) |
|---|---|---|---|
| Primary AI Strategy | Hybrid (Internal/DeepSeek/Nvidia) | Vertical Integration (FSD/Dojo) | OpenAI Partnership |
| Compute Platform | Jetson Thor | Custom FSD Chip | Custom/Nvidia Hybrid |
| Market Focus | Industrial/Research | Mass Consumer/Manufacturing | Commercial/Logistics |
| Model Architecture | WMA/VLA | End-to-End Neural Net | Multimodal VLA |
🛠️ Technical Deep Dive
- H2 Plus Architecture: Integrates a dual-stream processing pipeline where the WMA handles long-term spatial planning while the VLA manages high-frequency motor control loops (1kHz+).
- Cosmos 3 Integration: Utilizes a sparse-attention mechanism to minimize memory footprint on the Jetson Thor, allowing for larger context windows in complex, unstructured environments.
- WMA Implementation: Employs a predictive coding framework that simulates physical outcomes before execution, reducing the need for constant sensor-to-actuator feedback loops.
- Jetson Thor Specs: Features a Blackwell-based GPU architecture providing up to 800 TFLOPS of FP8 compute for real-time embodied inference.
🔮 Future ImplicationsAI analysis grounded in cited sources
Unitree will achieve sub-50ms latency in complex manipulation tasks by Q4 2026.
The integration of Nvidia's Jetson Thor with the optimized Cosmos 3 model is specifically engineered to eliminate the bottleneck between perception and motor execution.
Unitree will pivot to a 'Robot-as-a-Service' (RaaS) model to offset declining hardware margins.
The significant R&D investment in AI 'brains' suggests a shift toward recurring software revenue rather than relying solely on one-time hardware sales.
⏳ Timeline
2023-08
Unitree releases the H1 humanoid robot, marking its entry into the general-purpose humanoid market.
2024-05
Launch of the Unitree G1, a lower-cost humanoid aimed at developer and research markets.
2025-11
Unitree completes its IPO on the Hong Kong Stock Exchange, securing capital for AI model development.
2026-03
Unitree announces the strategic partnership with DeepSeek to enhance embodied AI capabilities.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗