01
Product catalog cheat sheet2026-05-10
Big Robotics Companies
Core products and meeting angles
Hangzhou
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
04
Product catalog cheat sheet2026-05-10
05
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
12
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
14
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
16
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
18
Product catalog cheat sheet2026-05-10
19
Product catalog cheat sheet2026-05-10
Product catalog cheat sheet2026-05-10
AI Labs
Research groups and collaboration routes.
01
Shenzhen Institute of Artificial Intelligence and Robotics for Society
Research area
CUHK-Shenzhen-led AIRS is a Longgang institute for fundamental and applied AI + robotics, with research centers spanning Embodied AI, Robot Systems, Healthcare Robotics, Micro-Nanorobotics, and Crowd Intelligence. The Embodied AI Center connects AIRSHIP, an official open-source embodied robot software stack for natural-language navigation and grasping on wheeled-arm robots with LiDAR, camera/RGBD, gripper, and Nvidia Orin-class embedded compute; AIRSPEED, an open-source data-production platform for collection, generation, and dataset construction that explicitly balances latency, transmission bandwidth, data quality, and compute resources; OmniRL multi-embodiment policy work; and AIRSTONE, an accelerator project for real-time, power- and cost-constrained robot computing that reports FPGA/dataflow accelerator work including a 23% end-to-end latency reduction and 40x–60x factor-graph speedups over embedded ARM baselines.
- Readiness
- ready
- Location
- 14F, Tower G2, Xinghe World, Yabao Road, Longgang District, Shenzhen.
02
Tencent Robotics X Lab
Research area
Shenzhen Robotics X lab established by Tencent in 2018 for mechatronics, perception, and control research; Tencent Robotics X Team and Futian Laboratory's TAIROS paper frames TAIROS as a modular embodied-AI platform built on SLAP3, with Embodied Perception, Embodied Planning, and Perception-Action (PAction) modules for multimodal scene understanding, long-horizon task decomposition, locomotion, and manipulation. The paper says the system can run as a full agent or standalone services through API calls or SDK-based edge deployment, and reports validation across humanoids, quadrupeds, and bi-manual manipulators on commercial robot platforms including Unitree, PaXini, Leju, Dobot, and EngineAI.
- Readiness
- ready
- Location
- Tencent Binhai Building, Nanshan, Shenzhen.
03
XbotPark
Research area
Dongguan / Songshan Lake robotics and smart-hardware incubation platform founded by Professors Li Zexiang, Ping K. Ko, and Gan Jie. Official XbotPark pages position it as a global startup incubator for robotics and smart hardware, with R&D facilities, office and living space, supply-chain and XbotPark Collaborative Manufacture support, funding through XbotPark Fund and VC partners, university talent pipelines, channel partners, and public-policy support for hardware commercialization. Official portfolio pages list 80+ robotics and smart-hardware companies, six unicorns, and named examples including Narwal, ePropulsion, HAI ROBOTICS, CiDi, SwitchBot, and LiberLive.
- Readiness
- ready
- Location
- Songshan Lake, Dongguan; No.11 Daxue Road, Songshan Lake Zone, Dongguan, Guangdong, China.
04
Beijing Academy of Artificial Intelligence
Research area
Foundation-model research; embodied-AI platform work through E12 / Robo X Studio and RoboBrain materials. BAAI's official site lists embodied large models as a core research direction and its open technical stack spans FlagData, FlagOpen, FlagOS, and FlagEval; the public RoboBrain line adds a CVPR 2025 unified brain model for robotic manipulation, RoboBrain 2.0 3B/7B/32B variants for perception, reasoning, and planning, plus related ShareRobot / ShareRobot-Bench evaluation assets. RoboOS extends the stack into a brain-cerebellum framework for cross-embodiment and multi-agent coordination with real-time shared memory and real-world restaurant, household, and supermarket demos across single-arm, dual-arm, humanoid, and wheeled embodiments.
- Readiness
- ready
- Location
- Beijing; public embodied-intelligence routes through BAAI / EI2 and open RoboBrain-RoboOS project channels.
05
Shanghai Artificial Intelligence Laboratory
Research area
Foundation-model and embodied-AI platform work centered on Shanghai AI Lab's public research stack: official talent pages cover large models, multimodal models, embodied intelligence, safe/trustworthy AI, AI for Science, open AI platforms, basic software, and AI hardware; the Intern Robotics full-stack engine combines simulation, data, training/evaluation tools, general embodied foundation models, and hardware interfaces into a perception-imagination-execution loop. Public assets include InternVLA for multi-embodiment VLA work, InternUtopia simulation, InternNav, InternManip, InternHumanoid, InternWorldModel / 4D world-model work, a data engine with 2M+ open data across 11 datasets and LeRobot-style formats, plus OpenGVLab vision/VLM work such as InternVL, InternVideo, PointLLM, and OmniQuant.
- Readiness
- ready
- Location
- L1, International Media Port, 129 Longwen Road, Xuhui District, Shanghai.
06
Institute for AI Industry Research
Research area
Tsinghua AIR lab-to-industry institute for AI + industrial systems; DISCOVER lab is explicitly robotics. It covers digitization of physical systems, collaborative human-machine intelligent systems, multi-sensor collaborative perception, and multi-agent decision-making. AIR DISCOVER is led by Guyue Zhou, a Tsinghua associate professor with DJI robotics, computer-vision, and intelligent-manufacturing background, and the DISCOVERSE work uses 3D Gaussian Splatting plus MuJoCo and ROS2 for high-fidelity Real2Sim2Real robot learning across AIRBOT Play, AIRBOT MMK2, wheeled loco-manipulators, quadrotors, and other robot agents.
- Readiness
- ready
- Location
- 12/F, Block C, Qidi Science and Technology Building, Tsinghua Science and Technology Park, Haidian District, Beijing.
07
National and Local Co-built Humanoid Robotics Innovation Center
Research area
Shanghai/Pudong humanoid-robotics public platform built around Humanoid Robot (Shanghai) Co., Ltd.; MIIT awarded the national-local co-built center in May 2024 to focus on common-key technology R&D, public incubation, talent training, core think-tank work, and industrial public services. Its OpenLoong / QingLoong open community provides full-stack humanoid resources: open-source full-size hardware, control framework, whole-body dynamics control, algorithms, datasets, tools, documentation, forums, and community support. Public reporting describes the Zhangjiang heterogeneous humanoid training facility as 5,000+ square meters with 100+ robot types from 12+ enterprises, a phase-one library of about 45 atomic skills such as grasping, picking, placing, and transporting, and real-machine data output targets rising from 20,000-30,000 entries per day in testing to 50,000 per day at full operation.
- Readiness
- ready
- Location
- Zhangjiang Robot Valley / Pudong, Shanghai; OpenLoong community operated by Humanoid Robot (Shanghai) Co., Ltd.
VCs / investor / angel routes
Capital and ecosystem leverage, not generic fundraising.
01
Qiming Venture Partners
Investment focus
Top-tier China venture platform founded in 2006, investing early and growth across Technology and Healthcare. Official Qiming pages describe 11 USD funds, 7 RMB funds, $9.5B raised, 580+ portfolio companies, 210+ exits, and 80+ unicorn or super-unicorn companies; stated technology focus areas include AI, robotics, semiconductors, enterprise software, and manufacturing upgrade. The most relevant route is the Technology practice around Managing Partner Alex Zhou, whose official profile names AI and robotics investments including UBTECH Robotics, Roborock, Yunji Technology, Mech-Mind Robotics, Galbot, TARS, Zhipu AI, Biren, Axera, StepFun, and DeepWay; Qiming's broader pages also surface WeRide, Hesai, and UBTech as portfolio highlights.
- Readiness
- ready
- Location
- Shanghai / Beijing / Hong Kong / Suzhou / Singapore
02
HAX / SOSV
Investment focus
Hands-on hard-tech venture platform for pre-seed and seed founders building physical, science-heavy, and engineering-heavy companies. HAX/SOSV material describes up to $550K in pre-seed funding, a six-month residency, a 35,000-square-foot Newark facility with labs, machine shop, electronics, 3D printing, metal fabrication, CNC, welding, and metrology, plus engineering, design, sourcing, manufacturing, and fundraising support through Newark, Pune, Shenzhen, and Tokyo. The focus map explicitly includes robotics, automation, advanced manufacturing, industrial independence, sensor data, AI, IoT, climate, health, and supply-chain resilience; robotics proof includes Rightbot freight-unloading robots, Danu Robotics material-sorting systems, Namu Robotics tree-planting robots, Tensorfield Agriculture computer-vision weeding robots, and other HAX automation demo-day companies.
- Readiness
- ready
- Location
- Newark / Shenzhen / Pune
03
Lenovo Capital & Incubator Group
Investment focus
Lenovo's corporate-venture and incubator arm for the smart ecosystem: official Lenovo pages frame LCIG around IoT, edge computing, big data, cloud, AI, smart verticals, early-stage technology VC, and internal incubation. LCIG/Lenovo material cites 250+ technology investments, 18 IPOs, a RMB 20B AI-terminal ecosystem plan, 50+ AI-related companies, and a decade of AI+robotics investing across core components, machine vision, mobile/service robots, compound robots, and humanoids. Official LCIG stories name robotics and embodied-AI examples including LimX Dynamics, Cross-Dimensional Intelligence, Star Era / Robot Era, Benmo, Yunji Technology, Future Robot, and its Fire/Light programs for humanoid manufacturing-as-a-service.
- Readiness
- ready
- Location
- Beijing / Shenzhen
04
Li Zexiang via XbotPark
Investment focus
Founder-angel, incubator, and manufacturing-readiness route around Prof. Li Zexiang and XbotPark: public pages frame XbotPark as a robotics and smart-hardware incubator with 80+ portfolio companies, six unicorns, examples including Narwal, ePropulsion, HAI ROBOTICS, CiDi, SwitchBot, and LiberLive, plus XbotPark Fund, VC partners, R&D facilities, founder housing, and supply-chain help from design and prototyping through core components, logistics, and commissioning.
- Readiness
- ready
- Location
- Songshan Lake, Dongguan
05
Tencent
Investment focus
Strategic platform route rather than classic VC: Tencent Robotics X was established in Shenzhen in 2018 and researches mechatronics, perception, and control; its TAIROS embodied-AI platform packages multimodal perception, long-horizon planning, and perception-action modules across humanoids, quadrupeds, and bi-manual manipulators via APIs, SDKs, cloud services, and edge deployment.
- Readiness
- ready
- Location
- Shenzhen
Events + people
Calendar anchors and connector routes.
Events
101
World AI Device Expo 2026 & 7th Shenzhen Global AI Expo
May 14–16, 2026
Calendar anchor
May 14–16, 2026 · Hall 1, Shenzhen Convention and Exhibition Center, Futian District
- Date
- May 14–16, 2026
- Location
- Hall 1, Shenzhen Convention and Exhibition Center, Futian District
People routes
0No individual connector routes yet.
Add angels, family offices, founders, or power brokers when there is a plausible warm path and concrete ask.
Open decisions
- Which capital route should be warmed first: Qiming for market map, HAX/SOSV for cross-border hard-tech, Li Zexiang/XbotPark for ecosystem power, Lenovo for infrastructure, or Tencent for platform strategy?
- Which individual connector routes belong here: angels, founders, family offices, operator-investors, or ecosystem guides?
- Should the World AI Device Expo become the calendar anchor for Shenzhen meetings, or just a scouting stop between company visits?
- Which outreach mode is preferred for complete rows: founder email, BD contact, PR/media route, investor-introduction route, or warm ecosystem intro?
- Should the next pass add Chinese names, addresses, WeChat-friendly intro blurbs, owner, priority, and meeting-date columns?
Reflex pitch bridge
Bilingual cheat sheet for China investor rooms.
Show this card
Reflex gives consumer robots a cloud reflex arc.
Reflex serves ultra low latency, secure inference on the cloud for consumer robots: p50 19 ms / p95 22 ms secure inference per chunk in pick-and-place tests at ~30 Hz, with a robot in the HF0 house and an H100 in us-west-1.
Reflex 为消费级机器人提供云端“反射弧”：安全、超低延迟的模型推理，在约 30 Hz 抓取与放置测试中，每个分块达到 p50 19 毫秒 / p95 22 毫秒。
Median cloud-to-robot secure inference per chunk in the HF0 pick-and-place test.
Tail percentile from the same test; keep it as a number until p99, max, jitter, and sample size are known.
Pick-and-place tests with a robot in the HF0 house.
The policy was post-trained to stay comfortable with delay.
The post-training framework improved pi0.5 iteration speed.
Cloud-side secure inference worked cleanly at this batch size.
Data needed before charting
Keep this numeric until the latency dataset is real.
- sample sizeChunk count, run duration, number of trials, task success rate, and environment notes.
- p99 / max latencyOutlier tail, jitter, retries, drops, and whether rare spikes break manipulation.
- network pathRobot location, cloud region, RTT split, secure-channel overhead, and inference-only timing.
- benchmark controlsModel, hardware, batch settings, measurement timestamp, and whether security overhead is included.
Why it matters
Cloud intelligence without robot-side GPU bloat.
The pitch is not generic robotics SaaS. It is a latency, security, and post-training stack that lets consumer robots call bigger cloud models while still acting inside the physical control loop.
Training claim
Delay is part of the policy, not an afterthought.
The post-training framework trains pi0.5 roughly 2× faster and was used to train with 40 ms lag in mind; the result worked well for batch-32 inference.
The simple term for model inference fast enough to matter for robot response.
Use this for the moat: cloud inference that behaves like a reflex, not a slow API call.
Signals privacy, IP protection, and controlled cloud execution.
Good for home robots, companion devices, desktop arms, and service robots sold at scale.
Makes the cloud part legible without sounding like generic SaaS.
A familiar manipulation benchmark: grab an object and put it where it belongs.
Frames the pi0.5 improvement as an engineering platform, not a one-off run.
Explains why a policy can remain stable when the cloud path has real network lag.
Rapport opener: humble, specific, and useful in investor or founder rooms.
Plain-language investor bridge for the latency chart.