On September 22, 2026, three independent pieces of progress arrived in embodied intelligence on the same day. Li Auto's Foundation Model team released the world-action-tactile model ME-Dex-1.0, upgrading touch from an "auxiliary condition" to a future world state the model must predict, with a tactile autoencoder jointly denoising across video, action, and touch experts. Guangdong Province, at a "AI+" manufacturing matchmaking conference, announced results from its "Thousand Robots into Factories" scale-up program: over 2,000 embodied robots have entered industrial scenarios province-wide, with 8 projects signed on-site totaling 75 million RMB. XPeng held its first robot supply chain partner conference the same day, announcing that core component designation agreements for the IRON humanoid have been signed and mass production scheduling is underway.
The three tracks are independent, but together they sketch where the industry stands: the model layer is learning "force", the industry layer is running "volume", and the supply chain layer is locking in "parts".
Li Auto ME-Dex-1.0: Touch Moves from Auxiliary to Target
The mainstream embodied-model approach treats video as the world state: the model predicts the next frame and generates actions from it. Cameras can't distinguish "smoothly inserted" from "jamming against an edge" — tactile signals give a clear verdict. Li Auto upgrades touch into something the model must predict, filling exactly this gap.
ME-Dex-1.0's technical path has three pieces.
First, a standard hand model. Two-finger grippers and five-finger dexterous hands differ in shape, sensor count, and placement; the same reading means something completely different on a fingertip versus a palm. ME-Dex maps tactile measurements from different hardware onto 34 bilateral hand regions: five-finger hands map to finger and palm regions, two-finger grippers map to thumb and index regions. Positions without sensors are explicitly flagged, avoiding mistaking "no measurement" for "measured zero contact".
Second, a unified tactile autoencoder. It compresses signals from different layouts into a shareable internal representation while preserving contact location, force magnitude, and contact changes. This turns heterogeneous sensor data into a unified input.
Third, three experts + H-Bridge. Video, tactile, and action experts each handle prediction, coordinated through H-Bridge shared attention (a mechanism that exchanges information across modalities at intermediate network layers). The action expert can use intermediate representations of future vision and future contact states during generation.
The data scarcity problem is handled by a tactile data engine: the team replays existing robot trajectories in simulation, records touch via simulated force sensors, and time-aligns it with vision and actions. The result is simulated contact data, letting datasets that lacked touch participate in joint training.
What the Ablations Reveal
Touch isn't "add it and it works." Li Auto's ablation experiments show success rates climbing step by step:
| Task | Official Baseline | ME-Dex-1.0 | Gain | |---|---|---|---| | RoboTwin Clean-only avg. success rate | 71.3% | 78.9% | +7.6 pp | | DexJoCo dexterous hand, 11 tasks | 56.4% | 65.3% | +8.9 pp | | ManiFeel insert/assembly avg. | 55.0% | 70.0% | +15.0 pp | | ManiFeel power plug insertion | 58% | 88% | +30 pp |
Power plug insertion jumping from 58% to 88% is the most striking figure in the data. Plug insertion is the classic "can't see the difference but can feel it" scenario: cameras can't distinguish smooth insertion from jamming at an edge, but force signals give a clear verdict. ME-Dex-1.0's largest gains in exactly these scenarios match the expectation that tactile information contributes the most there.
Guangdong's "Thousand Robots into Factories": 2,000 Units and 8 New Signings
Guangdong's Department of Industry and Information Technology announced at the September 22 Guangzhou matchmaking event: over 2,000 embodied robots have entered industrial scenarios province-wide, covering key processes like industrial sorting, loading/unloading, material transfer, equipment inspection, and collaborative work.
The 8 signed projects span industrial manufacturing, warehouse logistics, metro, community parks, sports, and smart housing — both production-line intelligence upgrades and urban service digitalization. Once fully deployed, they are expected to bring 600+ more robots into frontline scenarios.
More strategically significant is the "pairing of Guangdong provincial SOEs with the province's ten humanoid robot 'champions'": ten whole-machine companies — LimX Dynamics, Autoho, ZhiPingFang, Leju, EngineAI, Midea, XPeng, Honor, UBTech, and Dobot — are paired one-on-one with provincial SOEs, letting scenario "demand" pull technology "progress". This "SOEs provide scenarios, robot companies provide capability" binding is worth more than simple purchase orders: scenarios are embodied AI's scarcest resource, and only a stable scenario can start the data flywheel.
Guangdong's humanoid robot deployments in factories are expected to exceed 3,000 units for 2026. The 2,000-unit scale isn't staggering against national capacity, but the direction matters more than the number: repetitive, fault-tolerant processes like industrial sorting and inspection are the easiest places to close the commercial loop.
XPeng IRON: 76 DOF, 2,250 TOPS, September Supply Chain Designations
XPeng Robotics held its first robot supply chain partner conference on September 22, with top suppliers of joint modules, actuators, and precision structural parts attending.
IRON's core specs are set: 76 whole-body degrees of freedom, 21 DOF per hand, three in-house Turing AI chips delivering 2,250 TOPS of effective compute, supporting on-device deployment of physical AI large models — autonomous completion of complex tasks without remote operation. A 9-microphone array for sound-source localization plus visual lip-reading multi-modal coordination lets IRON autonomously identify the primary interlocutor in a three-person conversation.
The commercialization timeline is clear: scaled mass production at the end of 2026, official launch and delivery to Chinese and overseas markets in 2027, with monthly capacity ramping to thousands of units. On September 8, XPeng's robot production line officially went live, with the first high-level general-purpose humanoid robot completing automated final assembly and walking off the line autonomously.
On August 24, XPeng's robotics business completed a first funding round of over US$900 million at a post-money valuation above US$6.3 billion — a record single private round for China's embodied AI industry. IDG Capital led, Gaorong Ventures participated, with Tencent and Alibaba joining as strategic investors. Taken together — the funding round, supply chain designations, production line launch, and end-of-month mass production — the pace is tighter than most peers.
Putting It Together: Model, Scale, Supply Chain — Three at Once
| Dimension | Li Auto ME-Dex-1.0 | Guangdong "Thousand Robots" | XPeng IRON | |---|---|---|---| | What it solves | Model layer lacks touch | Industry layer lacks real scenarios | Supply chain lacks stable mass production | | Key number | Power plug 58%→88% | 2,000 units deployed + 8 projects signed | 76 DOF / 2,250 TOPS / thousands of units per month | | Maturity | Academic benchmarks | Commercial contracts | Production readiness |
The three are not substitutes: the model is learning a new modality, the industry is running new scenarios, the supply chain is locking new parts. For embodied AI to move from "demo videos" to "reliably working", all three must advance in parallel.
The road actually running today is still a narrow one. International Federation of Robotics (IFR) data offers a comparison: roughly 7,000 humanoid robots were sold globally in industrial and professional-service segments in 2025, versus about 542,000 newly installed conventional industrial robots the same year. Bank of America projects humanoid shipments could reach 90,000 units this year and 1.2 million by 2030. Ninety thousand versus 540 thousand — about 17%. Moving humanoids from "demos" to "general-purpose" will take more days like September 22, where everything advances together.