I spent a day at a robot “carnival” in Shanghai. Here’s what I saw.

【Executive Summary & Core Event】

Humanoid robots are experiencing a dramatic surge in visibility and capability across China, with Shanghai serving as a prominent showcase city for what the country's leadership envisions as the next frontier of artificial intelligence: embodied AI. The MIT Technology Review's on-the-ground reporting from a robot 'carnival' in Shanghai reveals a landscape where humanoid machines are no longer confined to laboratory demonstrations but are being presented as viable participants in everyday economic and social activities. This movement is not incidental but represents a deliberate, state-coordinated strategy embedded in China's latest five-year plan, which explicitly identifies embodied AI as a critical technology pillar for national competitiveness and economic transformation.

The event in Shanghai brought together a diverse ecosystem of Chinese robotics companies, including established players like UBTECH Robotics, Fourier Intelligence, and Agibot, alongside newer entrants backed by tech giants such as Xiaomi, Huawei, and Baidu. These organizations showcased humanoid platforms capable of performing tasks ranging from basic locomotion and object manipulation to complex multi-step workflows requiring real-time environmental understanding. The convergence of large language models (LLMs), vision-language-action (VLA) models, and sophisticated robotic hardware represents a paradigm shift from traditional teleoperated or pre-programmed robots toward systems that can perceive, reason, and act autonomously in unstructured environments. The Chinese government's strategic framing of embodied AI as essential infrastructure positions this sector for substantial state investment, regulatory support, and industrial policy coordination that mirrors earlier successes in areas like electric vehicles and 5G telecommunications.

Key companies on display demonstrated humanoids with varying degrees of sophistication. UBTECH's Walker series, now in its fourth generation, demonstrated improved dexterity with 20+ degrees of freedom in each hand and the ability to perform tasks like folding laundry and preparing beverages. Fourier Intelligence's GR-2 platform emphasized dynamic locomotion, including the ability to recover from pushes and navigate uneven terrain. Meanwhile, startups like Agibot showcased cost-reduction strategies aimed at bringing humanoid unit prices below $30,000, a threshold many analysts consider critical for commercial viability. The event also highlighted the integration of Chinese domestic AI models—including Baidu's ERNIE Bot, Alibaba's Qwen, and SenseTime's multimodal systems—as the cognitive backbone for these physical agents, reflecting China's broader push for technological self-sufficiency in AI infrastructure.

【Technical Architecture & Key Innovations】

The technical architecture underlying these Chinese humanoid platforms represents a significant evolution from earlier generations of industrial and service robots. At the core of the new wave is the integration of large language models and multimodal foundation models as the reasoning and planning layer for robotic agents. Rather than relying solely on traditional control systems built on finite state machines or behavior trees, these humanoids employ hierarchical architectures where a high-level LLM or VLA model generates task plans and natural language instructions, which are then translated into low-level motor commands by specialized control modules. This approach, sometimes called 'language-conditioned robotics' or 'foundation model-driven robotics,' allows the same underlying AI model to generalize across thousands of different tasks without requiring task-specific reprogramming.

The vision-language-action (VLA) paradigm is particularly significant in this context. Companies like Agibot and Fourier have developed proprietary VLA models trained on massive datasets combining visual observations, natural language instructions, and corresponding robotic actions. These models are trained using techniques including imitation learning from human demonstrations, reinforcement learning from human feedback (RLHF), and increasingly, reinforcement learning in simulation environments. The training data pipelines often incorporate millions of hours of teleoperated robot data, where human operators remotely control robots through VR headsets or motion capture suits, generating the action-conditioned datasets that form the foundation for autonomous behavior. Some Chinese labs have reported training datasets exceeding 10 million action sequences, rivaling the scale of data used in frontier LLM development.

On the hardware side, these humanoids employ sophisticated sensor suites including RGB-D cameras, LiDAR, IMUs, force-torque sensors at joints, and increasingly, tactile sensing arrays on fingertips. The actuation systems have shifted toward electric motors with harmonic drives and planetary gearboxes, moving away from hydraulic systems that were common in earlier humanoid designs. Control frequencies of 100-1000 Hz enable the rapid feedback loops necessary for dynamic balance and manipulation. The integration of edge computing modules running optimized inference engines for on-device AI processing—often leveraging custom NPUs from Chinese chipmakers like Cambricon or Horizon Robotics—reduces latency and enables real-time decision-making without constant cloud connectivity. Several platforms demonstrated end-to-end latency from visual perception to motor actuation under 100 milliseconds, a threshold considered necessary for safe and responsive physical interaction in human environments.

【Industry Context & Competitive Landscape】

China's humanoid robot sector now operates within a fiercely competitive global landscape that includes OpenAI's Figure AI partnership, Tesla's Optimus program, Anthropic's ongoing research into embodied agents, and Google DeepMind's work on RT-2 and related multimodal systems. However, China's approach differs fundamentally from its Western counterparts in several key respects. The Chinese strategy emphasizes rapid industrialization and cost reduction over incremental capability improvements, with explicit government targets for mass production volumes and unit price reduction. This industrial policy approach mirrors China's playbook in solar panels and electric vehicles, where coordinated state support, domestic supply chain development, and aggressive pricing strategies enabled global market dominance within a decade. Several Chinese humanoid manufacturers have publicly committed to producing 100,000+ units annually by 2025, a scale that dwarfs current Western production plans.

The competitive dynamics are further shaped by China's advantage in manufacturing infrastructure and supply chain integration. Chinese humanoid manufacturers benefit from proximity to the world's most advanced robotics component supply chains, including precision gearboxes, harmonic drives, and servo motors, many of which are produced domestically at costs significantly below Western equivalents. Companies like UBTECH have achieved vertical integration that includes in-house development of core components such as actuators and control boards, further reducing costs and dependencies. This manufacturing advantage allows Chinese firms to offer humanoids at price points that Western competitors struggle to match—some platforms are now priced below $20,000, compared to Tesla's estimated $30,000 for Optimus and Figure AI's undisclosed but reportedly higher price points.

In terms of AI model capabilities, Chinese companies are leveraging both domestic foundation models and open-source architectures to power their humanoid platforms. Baidu's ERNIE Bot, Alibaba's Qwen series, and DeepSeek's models all serve as reasoning backbones for various humanoid systems, while companies also integrate open-source models like Llama and Mistral where appropriate. The Chinese ecosystem's access to massive real-world data from its dense urban environments, extensive manufacturing sector, and large service economy provides unique training advantages for embodied AI systems. However, Western companies maintain advantages in frontier model capabilities, particularly in reasoning and generalization, and in access to advanced semiconductor manufacturing for training hardware. The race is thus multidimensional, spanning AI model quality, hardware engineering, manufacturing scale, and deployment ecosystem development.

【Developer & Enterprise Implications】

For developers and enterprises considering humanoid robot deployment, the current state of Chinese platforms presents both significant opportunities and substantial challenges. On the opportunity side, the rapidly declining cost curve and improving capability profiles make humanoids increasingly viable for specific use cases, particularly in manufacturing, logistics, and service environments. Chinese factories have already begun pilot deployments of humanoids for tasks like material handling, quality inspection, and assembly line assistance. The integration of domestic AI models means that Chinese enterprises can deploy these systems without dependency on foreign AI infrastructure, an important consideration given ongoing US-China technology restrictions and data sovereignty requirements.

Integration complexity remains a significant barrier to widespread adoption. Deploying a humanoid robot in a production environment requires not only the physical unit but also supporting infrastructure including charging stations, safety systems, network connectivity, and integration with existing enterprise software and control systems. Most Chinese humanoid platforms currently offer SDKs and APIs for custom task development, but the maturity of these developer ecosystems lags behind established robotics platforms like ROS (Robot Operating System). Companies interested in deployment must invest in specialized engineering talent capable of bridging traditional robotics engineering, AI/ML development, and domain-specific process knowledge. The typical deployment timeline for a humanoid robot in a manufacturing setting currently ranges from 3-6 months from initial evaluation to full operational deployment.

Hardware requirements and deployment costs vary significantly across the Chinese humanoid market. Entry-level platforms suitable for basic demonstration and limited task execution are available at $10,000-20,000, while production-grade units with full dexterity, advanced perception, and extended battery life range from $30,000-100,000. Battery technology remains a constraint, with most platforms offering 2-4 hours of continuous operation before requiring recharge. The business case for humanoid deployment currently centers on labor-cost arbitrage in environments where human labor is expensive or scarce, repetitive tasks with high consistency requirements, and hazardous environments unsuitable for human workers. Chinese companies are actively developing business models including robot-as-a-service (RaaS) subscriptions that lower the upfront capital barrier for enterprise adoption.

【Key Takeaways & Strategic Outlook】

The Shanghai robot carnival represents a pivotal moment in the global AI and robotics landscape, signaling China's serious commitment to embodied AI as a strategic technology domain. The convergence of increasingly capable foundation models, rapidly improving robotic hardware, and aggressive industrial policy creates a trajectory where humanoid robots could transition from novelty to necessity within the next 3-5 years. The key strategic insight is that embodied AI represents a fundamentally different value proposition from software-only AI: it extends AI's reach into the physical world, potentially automating the vast majority of economic activity that currently requires human physical presence. This makes it one of the highest-stakes technology domains of the coming decade.

The competitive race in embodied AI will likely be decided not by any single technological breakthrough but by the ability to build complete, integrated systems at scale. China's advantages in manufacturing infrastructure, supply chain integration, and state-coordinated industrial policy position it strongly in this dimension. However, the quality of underlying AI models, the sophistication of software ecosystems, and the ability to create compelling developer experiences remain areas where Western companies and open-source communities maintain significant advantages. The most likely outcome is a bifurcated global ecosystem with Chinese platforms dominating cost-sensitive manufacturing and logistics applications while Western platforms lead in complex, knowledge-intensive service applications.

Looking forward, the next generation of embodied AI systems will likely see breakthroughs in several key areas: world models that enable robots to simulate and predict physical outcomes before acting, multi-agent coordination systems where teams of robots collaborate on complex tasks, and continuous learning capabilities that allow robots to improve through real-world experience without catastrophic forgetting. China's massive deployment ambitions create a unique natural experiment in scaling embodied AI, generating real-world data and operational experience that will accelerate the entire field. Enterprises and developers should monitor this space closely, as the decisions made in the next 18-24 months about humanoid robot adoption could reshape labor markets, manufacturing economics, and the fundamental relationship between AI and the physical world.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.