WAIC 2026: Embodied AI Shifts from Showmanship to Practicality, Multiple Routes Advance Commercialization

The 2026 World Artificial Intelligence Conference (WAIC) opened in Shanghai on July 17, with embodied AI becoming one of the most watched sectors. Over 200 companies exhibited, showcasing the full chain from core components, dexterous hands to complete robots and large models. Unlike previous years that emphasized backflips and dancing, this year's conference focused on robots' real-world deployment capabilities, with many companies demonstrating practical applications in industrial production lines, warehouse logistics, and home services.

Simulation-based and Data-driven Routes Parallel

Sudo Technology (valued at 20 billion yuan) adheres to a simulation-first technical route, debuting at WAIC with four demos: precision assembly, flexible packaging, mobile manipulation, and battery module production line. CEO Han Zheng stated that their Sim2Real success rate has reached 99%+, and they dared to let visitors bring their own objects for blind testing. The company uses a virtual-real fusion data pipeline, primarily based on large-scale simulation data, fine-tuned with a small amount of real robot data.

Riemann Dynamics (a subsidiary of Kunlun Wanwei) takes a human video pre-training route. Its Riemann-1.0 model topped the RoboCasa-365 benchmark with 62.6%, 8.4 percentage points higher than the previous SOTA. The model uses 232,000 hours of training data, of which 200,000 hours are first-person human videos, converted into actionable robot training materials via a proprietary data processing pipeline.

Small Models and Open Source Trends

Facial Intelligence open-sourced the MiniCPM-Robot series, including a 1.5B-parameter VLA model and a 0.9B-parameter tracking and navigation model. MiniCPM-RobotManip scored 53.5 on the RMBench memory test, far exceeding π0.5's 10.4; single-step decision latency is only 120ms, about half of π0.5's. MiniCPM-RobotTrack achieves fully local, offline deployment on Unitree Go2 Edu with end-to-end response latency of about 180ms.

Qizhi Openmind (under EFORT) proposed an "Android for robotics" vision, releasing the HumanGPT world model and a universal technology base including Openmind OS, HALO skill suit, Dayan data platform, and Modou IDE. It has already connected with leading manufacturers such as SIASUN, AUBO, and Leju.

Breakthroughs in Dexterous Hands and Core Components

Wanna Robotics showcased the Std16A industrial dexterous hand (16 active DOF, grip force ≥60N) and the Eco12 lightweight dexterous hand. Its micro servo electric cylinder weighs as little as 31g and delivers push-pull force up to 250N. The company's dexterous hands have been deployed in a brand's 618 real warehouse sorting operations.

RoboScience released the world's first cloud-based embodied large model, Visics, enabling "one brain controlling multiple hands" — after swapping hands, the system can resume work within 30 seconds without retraining. The model adopts a VLOA architecture, using object 3D point cloud trajectories as core tokens, and supports over 10 different brands of dexterous hands.

Long-horizon Tasks and Multi-robot Collaboration in Focus

Qianxun Intelligence's Moz1 demonstrated a "tidy living room" long-horizon task at the booth, autonomously handling environmental changes like a suddenly inserted paper ball. Its Spirit v1.6 adopts a VLA and world model integrated architecture. Moz1 has been deployed on CATL's production line for high-voltage test plug insertion, with a success rate of over 99%.

Yuanli Jiji, in collaboration with StepFun, used 6 robots to assemble 80,000 building blocks into a 3.5-meter Great Wall model in 15 hours, entirely driven by the DM0.5 model, showcasing multi-agent coordination and sub-millimeter precision control.

Xingchen Intelligence released the Lumo-2 implicit world-action model, demonstrating over 20 home task videos in one go, including frying eggs, flipping pancakes, weighing millet, and catching rolling balls. End-to-end inference speed is 2.71 times faster than standard autoregressive models. It also launched the agent Philia, supporting persistent semantic memory.

Evaluation Systems and Data Infrastructure

RoboDojo released a unified simulation and real robot evaluation system, including 42 simulation tasks and 18 real-world tasks. Evaluations show that the best current general policy achieves an average success rate of only 8.80% in simulation and 12.8% in the real world, while human experts achieve 76.03% and 100%, respectively, indicating significant room for improvement.

JD.com released the JoyAI series models and the EgoLive dataset (2000 hours of first-person video), proposing a three-stage roadmap: "digital intelligence - possession intelligence - embodied intelligence," and setting a goal to build the world's largest physical world operations center.

Industry Consensus and Challenges

Based on multiple exhibits, the embodied AI industry is forming several consensuses: simulation and real data need to be used in combination; small models can achieve performance comparable to large models through extreme compression; "one brain, multiple forms" (a universal brain adapting to different bodies) is becoming the mainstream architecture; long-horizon tasks and open-world generalization remain the main bottlenecks.

A Frost & Sullivan report shows that Pudu Robotics ranks first globally in commercial service robot revenue and shipments, with over 130,000 cumulative deployments, generating 36.5 million hours of navigation data and 15.8 million hours of manipulation data annually, validating the value of large-scale deployment for the data flywheel.