WAIC 2026: Embodied Intelligence Breakthroughs from Household to Industrial Robots

The 2026 World Artificial Intelligence Conference (WAIC) opened in Shanghai on July 17, with embodied intelligence as the core focus. Multiple companies showcased full-chain progress from foundation models and dexterous hands to real-world deployment, marking the industry's shift from single-task demos to long-horizon, generalizable, and deliverable industrialization.

Model Competition: VLA Meets World Models, Memory and Generalization Key

Several companies released next-generation embodied foundation models, with breakthroughs in long-horizon task memory, cross-embodiment generalization, and data efficiency.

• ModelBest open-sourced the MiniCPM-Robot series, including the general VLA model MiniCPM-RobotManip and the tracking/navigation model MiniCPM-RobotTrack. The former achieves 53.5 points on the RMBench memory benchmark, far ahead of pi0.5's 10.4 points, with single-step decisions in 120ms; the latter is the industry's first purely local, offline tracking model with a tracking rate of 89.81%.
• Astribot released Lumo-2, the industry's first implicit world-action model for home scenarios, using three-stage progressive cross-modal alignment. It comprehensively outperforms pi0.5 and Fast-WAM in real-world complex task tests, with 2.71x faster end-to-end inference. It also launched the agent Philia with persistent semantic memory and modular architecture.
• Qianxun AI showcased Spirit v1.6, using a VLA and world model integrated architecture with abstract action token prediction. In the "tidy living room" long-horizon demo, it achieves dynamic replanning and anomaly handling. Its robot Moz1 performs high-voltage test plug insertion on CATL production lines with over 99% success rate.
• ForceMinds released the DM0.5 general foundation model, introducing a context abstraction layer (60-second memory), embodied CoT, and trajectory alignment layer. It achieves SOTA on RoboChallenge Table30 V2 with 43% success rate and 54.42 points. It also released the world model DW0.5-driven post-training framework DFOL2.0, ranking first on both EWMBench and WorldArena.
• GalaxyBot released the world's first embodied test-time post-training framework WAM-TTT, enabling robots to quickly adapt to new scenarios during deployment using human videos, without retraining or action annotations. In data ablation experiments, 100 robot trajectories + 100 human videos achieve 74.1% success rate, close to full robot data performance.
• RoboScience launched the world's first cloud-based embodied large model Visics, using a VLOA architecture with object trajectories as unified tokens, enabling "one brain controls multiple hands, 30-second hand swap". It supports over 10 dexterous hands with over 99% multi-SKU grasping success.
• Riemann Dynamics (Kunlun Tech subsidiary) released the Riemann-1.0 world action model, trained on 232,000 hours of data (including 200,000+ hours of human video). It tops the RoboCasa-365 household benchmark with 62.6%, an 8.4 percentage point improvement over SOTA.

Dexterous Hands and Actuators: From Display to Real Operations

Dexterous hands, as the terminal connecting cognition and the physical world, demonstrated capabilities from industrial sorting to fine manipulation at WAIC.

• Wanna Robotics exhibited the Std16A industrial dexterous hand (16 active DOF, grip force ≥60N) and the Eco12 lightweight anthropomorphic hand (12 DOF, 500g). Its micro servo electric cylinder weighs as little as 31g with push-pull force up to 250N. In May 2026, Wanna's dexterous hand entered a real warehouse for 618 shopping festival sorting operations, becoming one of the earliest to apply dexterous hands in real business scenarios.

Data and Evaluation: Standardization Accelerates

Data scarcity and inconsistent evaluation standards are long-standing industry pain points. Multiple works at WAIC addressed this.

• Alibaba DAMO Academy open-sourced the RynnWorld-Teleop digital teleoperation scheme, using generative world models to replace real robots. Operator gestures drive real-time video generation, automatically obtaining joint-level action labels. Real robot experiments show zero-shot Sim2Real transfer with synthetic data, and combining real data stably improves success rates.
• RoboDojo team released a unified evaluation benchmark covering simulation and real robot operations, including 42 simulation tasks and 18 real tasks. Evaluations show the best current general policy achieves only 8.80% average success rate in simulation and 12.8% in the real world, while human experts achieve 76.03% and 100% respectively, indicating huge room for improvement.

Extreme Challenge: Multi-Robot Collaboration and Long-Horizon Tasks

ForceMinds, together with StepFun, launched the "Great Wall Block Challenge" at WAIC: 6 robots (4 desktop + 2 wheeled) autonomously assembled a Great Wall model (3.5m×1.5m×1.1m) from 80,000 micro blocks within 15 hours, with smallest component precision of 0.1-1mm. No remote control or preset scripts; fully driven by the DM0.5 model, demonstrating sub-millimeter fine manipulation, multi-agent collaboration, and long-horizon stable operation.

Industry Trends: From Demo to Delivery, Core Components and Scenario Deployment

WAIC reveals three major trends in embodied intelligence:

• Model Architecture Fusion: VLA and world models move from competition to hybrid architectures; next-generation robot foundation models will likely combine both.
• Data Efficiency Improvement: Low-cost data sources like human videos, simulation data, and digital teleoperation are gradually replacing expensive real robot teleoperation data, lowering scaling barriers.
• Real-World Validation: Dexterous hands enter warehouse production lines, robots enter power battery lines. Industry evaluation shifts from "can it complete one action" to "can it consistently complete full tasks".

According to the "2026 Robot Industry Panorama Research Report", China's robot market is expected to grow from 200.3 billion RMB in 2025 to 491.4 billion RMB in 2030, a CAGR of 19.7%. In humanoid robot BOM costs, actuators account for about 40%-50%, and dexterous hands about 14%-18%.