Concluded | The Symposium on Embodied Perception Fusion & Multimodal LargeModel inovation

Gasgoo Munich-As embodied intelligence accelerates from "algorithmic concept" to physical interaction, the fusion of multimodal perception and large models is emerging as a critical path for industrial adoption. From VLA architectures reshaping decision-making paradigms to VLMs pushing machines from "seeing" to "understanding," and the integration of vision, touch, and force sensors, bridging the "perception-cognition-action" loop has become the core proposition for the next stage of development. Yet, practical challenges—computing power, data, on-device deployment, and hardware-software compatibility—are putting the transition from laboratory to market to the test.

Against this backdrop, Gasgoo Embodied Intelligence hosted the "The Symposium on Embodied Perception Fusion & Multimodal LargeModel nnovation 2026" in Shanghai on September 15. Experts from large models, robotics, chips, and sensors gathered to discuss hot topics ranging from technical paradigms and multimodal fusion to VLA architectures, cross-modal representation, computing power, and edge-cloud collaboration. The dialogue focused on the critical path from "algorithm" to "entity" and the industrial practice of fusing multimodal perception with embodied large models.

Host Welcome Remarks

Gasgoo Embodied Intelligence CEO Gu Xiaoying noted that the industry is at a pivotal stage, shifting from exploration to scale. The sector must tackle stability, generalization speed, cost, and safety. Positioning itself as a "connector and enabler" for the robotics industry, Gasgoo leverages its automotive supply chain resources to accelerate upstream and downstream connections. It has built an embodied intelligence database with over 2,000 robot-related companies, aiming to expand to more than 10,000 by year-end.

Building on this, Gasgoo is developing a service ecosystem covering supply-demand matching, content, data, events, and investment. By connecting industrial, healthcare, and educational scenarios with robot manufacturers, it aims to match technology with demand. Gasgoo also advocates for data co-creation, opening its database to companies and encouraging the sharing of non-confidential sales data to build a credible, globally perspective-driven system.

Gu Xiaoying | CEO, Gasgoo Embodied Intelligence

2026 WAIC & WRC Frontier Insights

Wang Xianbin, Gasgoo partner and VP of the research institute, analyzed the industry's progress based on WAIC and WRC findings. He projected global humanoid robot shipments at around 60,000 units this year, with China accounting for about 50,000. However, the sector remains in a transition phase from R&D to industrialization. Domestic applications are still dominated by entertainment and research, though deployment in 3C, automotive, and logistics has risen to 15%–20%. Some products are being tested in 4S stores, home environments, and pharmacies, but mass delivery is still a way off.

On the supply chain side, actuation systems account for nearly half the cost of humanoid robots. Given the high overlap with automotive supply chains, core components still offer significant room for cost reduction as production scales. Wang believes humanoids will gradually expand from structured, simple scenarios to unstructured ones, but small-scale pilots will remain the norm for the foreseeable future.

Technologically, Wang analyzed trends in large models, computing power, world models, and data systems. He noted that general-purpose LLMs are rapidly evolving in long-horizon task planning, reshaping industry value and creating opportunities for robot manufacturers and perception algorithm firms. Yet, as hardware and data capabilities advance, the industry faces a long cycle of verification and scenario expansion. True mass adoption depends on coordinated breakthroughs in perception, decision-making, execution, and cost control.

Wang Xianbin | Partner and VP of Research Institute, Gasgoo

Practicing Physical AI in Factories: From Simulation to Real-World Deployment

Unitree is building an AI-assisted closed-loop iteration system to drive the development of physical AI. The company remains committed to full-stack R&D, innovation, and scalable manufacturing for humanoid and quadruped robots, working with global developers to build a robust ecosystem.

Xie Yipeng | Head of Industrial Scenario Solutions, Unitree Technology Co., Ltd.

World Models: Opening a New Era of Physical AGI

Mao Jiming, partner and VP at GIGAAI, shared insights on general world models. Founded in 2023, JiJia focuses on building a technical system where the model's understanding and prediction of the physical world translate into action. Their framework comprises "world generation," which predicts changes from actions, and "world action," which generates actions based on observations. The roadmap moves from video understanding to spatial geometry and physical modeling, eventually reaching continuous real-world interaction.

JiJia has established a system combining algorithms and data, releasing its latest product in August. Its models have achieved continuous multi-task operation in home and industrial settings, validating the correlation between data scale and model performance. Mao believes that as world models improve their understanding and agency, they will drive mass adoption in robotics and autonomous driving.

Mao Jiming | Partner & VP, GIGAAI

Toward Ubiquitous Embodied Intelligence: Physically Native World Models, Data Structures, and Training Algorithms

Lu Yao, chief scientist at Sunrising AI Lab, outlined the shift from information-centric AI to physically interactive AI. Embodied intelligence, he argued, relies on a continuous loop of perception, decision-making, and control. Guangxiang proposes a "physically native" intelligence system designed at the model, data, and algorithm levels, combining explicit physics with neural networks and causal modeling to enhance real-world interaction.

The lab has built a one-stop development system covering data generation, training, and deployment. Its physically native world model is being validated in autonomous driving, humanoid robots (including continuous table tennis), and industrial robots for tasks like loading and quality inspection. Lu emphasized that moving AI into the physical world requires the co-design of models, data, and algorithms to evolve from single-scenario applications to general physical intelligence.

Lu Yao | Chief Scientist, Sunrising AI Lab

EngineAI's "Engine Plan" Gathers Ecosystem Forces to Drive Embodied Evolution

Zhao Feiyu, head of ecosystem at EngineAI, detailed the company's progress. Zhenqun maintains full-stack R&D across joints, whole machines, and control systems. Its products include the 173-centimeter T800 humanoid and the 135-centimeter PM01 lightweight agent. These are deployed in entertainment, commercial services, smart traffic, and manufacturing. Notably, they are working in Luxshare's Suzhou factory on material handling and with JD.com on store guidance.

On the ecosystem front, Zhenqun is opening its SDK and training frameworks to developers. The company plans to expand cooperation in industry co-creation, research, and talent development through initiatives like robot lending and project partnerships, aiming to accelerate the extension of robotic capabilities into diverse scenarios.

Zhao Feiyu | Head of Ecosystem, EngineAI

From Human Hand to Dexterous Hand: Tactile Gloves and Embodied Intelligence Data Collection

Liao Yongxing, AI algorithm director at Fulai New Materials, discussed tactile intelligence. While vision handles position and trajectory, tactile signals are crucial for grip adjustment and slip detection during contact. Tactile data significantly boosts success rates in fine manipulation. Liao argued that simulation, teleoperation, and "no-embodiment" data are complementary rather than substitutive.

Addressing issues like missing contact data and synchronization difficulties, Fulai has upgraded its tactile products—including gloves and dexterous hand terminals. Using technologies like multi-signal synchronization and drift compensation, the company improves data accuracy and consistency, laying the groundwork for training dexterous manipulation models.

Liao Yongxing | AI Algorithm Director, Fulai New Materials

From SLAM to World Models: Building the Cognitive Evolution Loop of Embodied Intelligence

Zhuo Weifeng, head of multi-sensor fusion at AiMOGA, introduced the company's product matrix. Founded in January 2025, Mojia offers robots for commercial, civil, and public services, including the Mo Ying humanoid in Chery's overseas 4S stores, hospital guide robots, and traffic police robots. It is also developing outdoor patrol dogs and a second-generation humanoid, Mo Qiong.

Technically, Mojia is moving from geometric mapping to semantic scene graphs and world models. By forming a data loop between geometric positioning, scene understanding, and world prediction, the system feeds prediction errors back to improve mapping and localization capabilities.

Applications are already live in the Yangtze River Delta. Traffic police robots interface with local public security data for violation detection and gesture commands, while home care robots facilitate real-time communication and hospital guide robots handle registration and consultation services.

Zhuo Weifeng | Head of Multi-Sensor Fusion, AiMOGA

Hotspot Dialogue: From "Seeing" to "Doing": Where Are the Key Breakthroughs for Embodied Intelligence to Reach General Intelligence?

The conference then moved to a panel discussion titled "From 'Seeing' to 'Doing': Where Are the Key Breakthroughs for Embodied Intelligence to Reach General Intelligence?" Hosted by JiJia's Mao Jiming, the panel featured Liao Yongxing from Fulai New Materials, Lu Yao from Sunrising AI Lab, Wang Guan from AiMOGA, and Ji Haifeng from RealMan Groupy.

The panelists debated whether VLA, world models, or physically native models constitute the "brain" of embodied intelligence. They also discussed how multimodal perception enters the decision loop, which data type—real machine, simulation, or human—will drive evolution, and the primary bottlenecks to scaling: models, data, perception, or hardware.

From Data to Capability: Mifeng's Million-Hour No-Embodiment Dataset and Full-Link Technical System

Liu Li, technical VP at Maniformer, presented the company's data infrastructure. Mifeng has built a million-hour dataset covering 500 tasks, 10,000 scenes, and 50,000 object types. Data is collected via bare hands, wrist devices, and physical equipment, integrating visual, tactile, force, and audio signals to bridge human intent and robotic replication.

Mifeng has established a data loop covering collection, processing, and feedback. Technologies like dynamic segmentation and pose reconstruction convert real-world actions into training data. By feeding execution errors back into the system, Mifeng is turning mistakes into model improvements, driving the transition from data collection to deployment.

Liu Li | Technical VP, Shanghai Maniformer Technology Co., Ltd.

Physical AI Driving Embodied Evolution: From High-Fidelity Simulation to Real-World Deployment

Shi Qianzi, eco-development director at Songying Technology, noted the shift from tech demos to actual delivery. Key constraints include data supply, multi-robot collaboration, and the lack of indigenous underlying tech. While real-machine data is accurate but costly, simulation pipelines are immature. Additionally, reliance on overseas tech stacks raises supply chain and compliance concerns.

Songying's solution is a proprietary Physical AI Operating System serving as a middleware layer. Compatible with nine domestic chip makers, it covers physics engines, rendering, and synthetic data generation. The system replicates real scenes with minimal disruption, creating a closed loop of scene collection, simulation, and deployment to turn tasks into reusable digital assets.

Shi Qianzi | Eco-Development Director, Songying Technology

Building a Data System for General Embodied Intelligence

Zhu Jie, VP at Kinetix AI, described the evolution of its "embodied brain" from fixed scenarios to humanoid migration. Early work on dual-arm control for non-rigid objects like clothing demonstrated strong generalization. The company is now exploring outdoor sports and ping-pong with humanoids, using a combined Vision-Action Model and World Model architecture to enhance semantic understanding and memory.

Zhu argued that traditional motion capture is too costly. Chaowei uses first-person human data for pre-training with lightweight headbands, enabling a "real-world pre-training plus fine-tuning" approach. He believes that general embodied intelligence relies not on a single architecture but on high-quality, growing real-world data to evolve robots from tools into general agents.

Zhu Jie | VP, Kinetix AI

Embodied Intelligence Data Training Grounds and the Ecosystem for Remote Operation Robots

Ji Haifeng, solutions director at RealMan Groupy, highlighted the high cost of data collection and the lack of standards. The focus is shifting to real-world deployment. He identified body, data, compute, and talent as core elements. While models drive entry, costs must drop from hundreds of thousands to tens of thousands of yuan for mass adoption.

Ruierman has built a product system based on wheeled robots and ultra-lightweight anthropomorphic arms. It adheres to data standards, collecting multi-dimensional data like images and calibration info to form a complete production pipeline.

Ji also introduced a remote operation model where humans control robots from offices to perform real production tasks. This allows continuous data acquisition without disrupting operations, gradually evolving from remote control to autonomous execution. The model is currently deployed in Changzhou.

Ji Haifeng | Solutions Director, RealMan Groupy (Beijing) Co., Ltd.

Embodied Intelligence Data Flywheel: Infrastructure Practice from Data Collection to Model Evolution

Liao Shaoyao, smart mobility solutions director at Tencent, noted that while real-world deployment is the consensus, data gaps and collection limitations remain challenges. Drawing parallels to autonomous driving, embodied AI will see rapid iteration and convergence, supported by terminal mass production and continuous data iteration.

Tencent's data flywheel spans collection, training, simulation, and deployment. It uses cloud channels for efficient remote operation data, a multimodal data lake for training, and cloud simulation to reduce costs. The company employs a cloud-edge-end architecture and continues to open its robot lab platform to foster the ecosystem.

Liao Shaoyao | Smart Mobility Solutions Director, Tencent

Embodied Data Driving Robot Intelligence Evolution: From Data Collection to Autonomous Learning Loop

Yan Wei, solutions expert at Wuwen AI, noted that embodied data is harder to acquire than autonomous driving data. To address high costs and toolchain gaps, Wuwen built a "collect, generate, simulate" loop. It has a 500,000-hour dataset in Deqing and conducts "wild collection" in factories and logistics, using no-embodiment data to lower barriers.

Wuwen's simulation toolchain supports asset building and training. It uses two generation paths: high-precision scanning and low-cost LLM-based generation. The company optimizes for complex objects like soft bodies and fluids and is building an evaluation system for basic actions like grasping. Yan believes a robust data-simulation loop is key to industry progress.

Yan Wei | Solutions Expert, Wuwen AI

Co-Evolution of Full-Function GPUs and Embodied Intelligence

Guo Hua, senior eco-director at Moore Threads, outlined three engineering pillars: model training, resource production, and agent production. Leveraging its full-function GPUs for AI, rendering, and physics, Moore Threads has trained world models and thousand-card embodied brain models, advancing domestic chip applications.

Moore Threads' workflow covers data generation, training, and deployment. It uses graphics rendering for data generation and domestic chips for VLA training. The company offers a robot simulation platform and the Yangtze River series SoCs for edge deployment. By linking cloud training, edge deployment, and data feedback, Moore Threads aims to provide a domestic computing foundation for the industry.

Guo Hua | Senior Eco-Director, Moore Threads

Soft-Hardware Integration and System Coordination: Building the AI Foundation for Embodied Intelligence

Chen Bendong, product marketing director at Qualcomm, traced AI's evolution from text and video processing to contextual and embodied applications. The core shift is moving from the information world to physical interaction, transforming AI from a tool into an autonomous agent.

The robotics industry is progressing from perception to dexterous operation and multi-agent collaboration. To meet demands for low power, low latency, and high intelligence, computing will form a continuum spanning terminals, edges, and the cloud, dynamically allocating tasks based on needs.

Qualcomm offers a chip matrix for various computing tiers, including the iQ series for industrial automation. The iQ-8 controls humanoid motion, the iQ-9 powers robot dogs, and the iQ-10 offers up to 700 TOPS—already proven in autonomous driving. The company collaborates with domestic humanoid firms, combining data loops with its software stack and leveraging experience from mobile and automotive sectors to support scaling in areas like thermal management and OTA updates.

Chen Bendong | Product Marketing Director, Qualcomm Technologies, Inc.

From Foundation Models to AOS: Engineering Practice of Edge-Cloud Collaborative Agents

Jiao Hui, solutions director at StepFun, discussed the transition from L3 to L4 models, where AI moves toward generating new methods and self-iteration. As AI expands from digital to physical realms via cockpits and glasses, StepFun has built an AOS architecture for edge-cloud collaboration. It balances intelligence, security, and latency, covering over 1 million vehicles in smart cockpits.

StepFun's model matrix spans service, lightweight, and edge models, extending into voice, image, and reasoning. Its end-to-end voice interaction has been validated at scale in cockpits. Jiao believes the AOS architecture will extend to AI glasses and embodied intelligence as edge computing improves, driving models into physical terminals.

Jiao Hui | Solutions Director, StepFun

AMD Physical AI Full-Stack Solution

Ying Yichen, physical AI market development manager at AMD, criticized the traditional "big and small brain" architecture where CPU handles motion and GPU handles perception, citing latency and waste. AMD promotes a unified memory fusion architecture, integrating motion control and AI to reduce latency from milliseconds to microseconds. Its new X100 (high-end) and P100 (mid-to-low) series offer up to 128GB unified memory and 50 TOPS NPU performance, with industrial-grade durability.

AMD provides a complete toolchain covering OS, AI components, and middleware. Using real-time Linux and virtualization, it meets deterministic latency requirements. A unified software stack allows model migration between cloud and edge. AMD also offers modular core boards and dev kits, supporting AI inference and multi-sensor access. The company plans to open reference designs for various robot types.

Ying Yichen | Physical AI Market Development Manager, AMD

Empowering Embodied Intelligence via Cloud-Edge-End Collaboration: From Single-Machine to Swarm Intelligence

Zhang Hui, chief robot product planner at ZTE, noted that robots have limited onboard compute due to space and power constraints, necessitating cloud-edge-end collaboration for dynamic resource scheduling. As robots move from single-machine to multi-agent intelligence, weakly coupled tasks can be cloud-scheduled, but strongly coupled tasks like dual lifting require millisecond-level sensor synchronization.

ZTE leverages its telecom and computing expertise for soft-hardware integration, developing industrial robots and dexterous hands while reusing technologies from its automotive sector. Solutions are deployed in commercial services and 3C manufacturing. Zhang plans to build a unified skill library to facilitate cross-robot reuse of models and protocols, accelerating the shift to swarm intelligence.

Zhang Hui | Chief Robot Product Planner, ZTE Corporation

New AI Infrastructure for Embodied Intelligence and Automotive: Chip-Cloud-Model Integrated Architecture and Strategy

Zheng Yang, senior AI automotive solutions architect at Alibaba Cloud, outlined the company's layout across chips, cloud, and models. It uses T-Head chips and heterogeneous computing for infrastructure, while the Qwen model supports open-source deployment from edge to cluster. As models shift from chat to task execution, the fusion of digital and physical worlds and multi-agent collaboration are becoming key trends.

In embodied intelligence, Alibaba Cloud has built a foundation system centered on the Qwen VLA model, with vertical models for QA, navigation, and operation. The operation model adapts to over 20 robot types, reducing migration costs. The navigation model handles target finding and obstacle avoidance, while the management model coordinates multi-robot tasks. Zheng emphasized the synergy between models, compute, and applications for scaling.

Zheng Yang | Senior AI Automotive Solutions Architect, Alibaba Cloud

The 2026 Embodied Perception Fusion and Multimodal Large Model Innovation Seminar concluded successfully. Guests exchanged views on tech evolution, perception, data, computing, and industrial landing, exploring the path from "seeing" to "doing." Gasgoo Auto will continue to monitor embodied intelligence, building platforms for exchange and cooperation to drive the fusion of innovation and practice.