2.1 The "Old Era" Crown (Approx. 2016-2021): The Glory of Model-Driven and Optimal Control

During this phase, the core philosophy in robotic dynamics was "omniscient model and precise control." Researchers pursued building an accurate physical dynamic model for the robot and designing control algorithms capable of solving for optimal motion trajectories in real-time.

• Core Logic: Model-based real-time optimal control. The core of the tech stack was Model Predictive Control (MPC), Whole-Body Control (WBC), and state estimation. Engineers needed to write extremely complex dynamic equations for the robot. The controller would calculate optimal joint torque commands within milliseconds based on the current state and the model to accomplish high-difficulty actions like dynamic balancing and jumping.
• Pinnacle Representative: The parkour and backflips of Boston Dynamics' hydraulically-driven Atlas. These demonstrations shocked the world, underpinned by top-tier hydraulic actuators, sophisticated sensor fusion algorithms, and a deep understanding and engineering realization of rigid-body dynamic equations. It demonstrated the performance limits achievable by traditional model-based control methods.

Technical Bottlenecks and Data:

1. Extreme Reliance on Precise Modeling: Any unmodeled disturbance (slippery ground, sudden load change) could cause system instability. As observed by the industry, such robots essentially "executed pre-set trajectories," lacking genuine understanding and adaptation to environmental changes.
2. Complex Code, Weak Generalization: Each new action required engineers to spend months on mathematical modeling, algorithm tuning, and parameter adjustment. The system's "intelligence" was entirely defined by human code, unable to autonomously handle new scenarios.
3. High Cost: Complex, noisy, and maintenance-intensive hydraulic systems meant they could only be expensive laboratory research platforms or special-purpose equipment, lacking the potential for large-scale commercialization.

Table 2-1: Representative Technology and Limitations of the Model-Driven Era (Taking Boston Dynamics' Hydraulic Atlas as an Example)

2.2 The "Transitional" Fusion (Approx. 2021-2024): The Symbiosis of Learning Algorithms and Traditional Control

With the rise of Deep Reinforcement Learning (RL), the paradigm relying solely on manual modeling began to loosen. Researchers started experimenting with data-driven methods to "generate" skills, but for safety and reliability, still used traditional controllers as the "safety valve" for execution.

• Core Logic: Learning generates "skills," control ensures "stability." A typical path: a "policy network" is trained using RL in a simulation environment. This network can generate desired motion trajectories or joint positions based on perceptual input. Then, this "rough" trajectory is fed into a model-based traditional controller (e.g., MPC), which is responsible for calculating precise torque commands and ensuring stable execution.
• Typical Representative: Learning and optimizing bipedal robot walking gaits. Many studies learned natural, energy-efficient walking patterns through RL in simulation, then deployed the learned policy to physical robots, with underlying WBC controllers ensuring they don't fall while walking.
• Significant Meaning: This marked the formal entry of data-driven methods into the core domain of robot control. It partially liberated engineers from manually programming every detailed action. However, this approach did not shake the underlying architecture centered on state estimation and optimal control; the learning part and the traditional control part remained relatively loosely coupled.

2.3 The "New Era" Paradigm Revolution (2024-2026): AI Models, World Models, and the Disruption of End-to-End Learning

Starting around 2024-2025, and especially marked by a series of announcements at CES 2026, the field of robotic dynamics has undergone a fundamental paradigm revolution. Its core driving force is the maturation of Physical AI, aiming to embed an "understanding" and "intuition" of the physical world into the robot's "brain."

• Core Logic: From "programmed behavior" to "emergent behavior," from "solving the known" to "coping with the unknown." Its technological pillars include:
  1. World Models (e.g., NVIDIA Cosmos): This is a learnable physics simulator that has learned to predict the evolution of the physical world through massive video data. It allows robots to conduct "thought experiments" before acting, predicting the consequences of different actions, thereby possessing near-human physical intuition.
  2. End-to-End Vision-Language-Action Models (e.g., NVIDIA GR00T): The model can directly generate continuous, whole-body robot action sequences from visual and language inputs. It bypasses the traditional modular pipeline of "perception-planning-control," achieving tight coupling of perception, reasoning, and decision-making.
  3. Bridging the Simulation-to-Reality Gap: Leveraging generative AI for domain adaptation allows strategies trained in simulation to transfer to real robots with near-zero-shot capability, solving a major obstacle that long hindered the deployment of robot learning algorithms.

Table 2-2: Comparison of Core Tech Stacks Between Old and New Paradigms (Early 2026 Perspective)

The Implication from the All-Electric Atlas

The all-electric Atlas did not abandon dynamics. On the contrary, through the engineering innovation of all-electric joint modules, it packaged the once-top-tier but cumbersome dynamics control capability into more quiet, reliable, and maintainable standardized modules. Simultaneously, it reserved interfaces for the AI "brain." It is foreseeable that future top robots will be a deep fusion of "AI models responsible for high-level task understanding and generalized planning" + "simplified and reinforced classical dynamics controllers responsible for underlying safety and stable execution." This paradigm revolution ignited by AI is redefining the connotation and boundaries of robotic dynamics.

3.1 First Pole: The Tech-Driven Camp – Boston Dynamics and the Industrial Breakthrough of "Body" Experts

This camp, with Boston Dynamics as its absolute core, represents a philosophy based on physical models and deep engineering. Their advantage lies in decades of accumulation in dynamic motion and dynamics control, possessing the highest level of technical completion and reliability validation.

However, as the release of the all-electric Atlas reveals, their strategy has undergone a fundamental shift: from pursuing "technical limits" to delivering "industrial value." To enter this scaled "Industrial Value" era, Boston Dynamics is no longer going it alone but has constructed an exemplary industrial ecosystem closed-loop:

Table 3-1: Boston Dynamics' Strategic Transformation and Ecosystem Building

This closed-loop of "Body + AI Brain + Compute + Scenario" signifies that the tech-driven camp is integrating into the mainstream industrial ecosystem with unprecedented openness, transforming top dynamics control capabilities into deliverable industrial productivity.

3.2 Second Pole: The AI & Data-Driven Camp – The Tech Giants' Battle for "Brains" and Platforms

Represented by tech giants like NVIDIA, Tesla, and Google, this camp enters from the cloud and algorithm layers, attempting to redefine the rules of the game. Their core logic is: to "brute-force crack" the robot control problem top-down with massive data, supercomputing power, and AI models.

NVIDIA's announcements at CES 2026 are a concentrated embodiment of this pole. It is no longer limited to providing chips but is building a complete robotic development "operating system" around the concept of "Physical AI":

1. Model Layer: The Cosmos World Model aims to give robots an understanding of physical laws, possessing "physical intuition"; the Isaac GR00T model is a Vision-Language-Action model that directly generates whole-body movements.
2. Toolchain: The open-source simulation platform Isaac Sim, and cooperation with Hugging Face, aim to connect global developers and establish standards.
3. Hardware Layer: The Jetson series edge computing platforms provide the computational power for deployment.

Tesla represents the path of vertical integration. Its Optimus robot completely serves its "data flywheel" strategy: acquiring data in the real world, training end-to-end models, and ultimately achieving large-scale, low-cost automation.

Table 3-2: Core Strategies and Representative Companies of the AI-Driven Camp (Early 2026)

Advantages and Bottlenecks: The AI-driven camp's advantage lies in the ultimate potential for intelligent generalization and strong ecosystem appeal. However, its bottlenecks are equally evident: the complexity and safety requirements of the physical world are extremely high; purely data-driven "black box" models carry unpredictable risks when dealing with edge cases; and efficiently and reliably deploying powerful cloud models onto resource-constrained robot bodies remains a major challenge.

3.3 Third Pole: The Product & Scenario-Driven Camp – The Agile Charge of Startups

Represented by startups like Figure AI, Agility Robotics, Unitree Robotics, and Fourier Intelligence, this camp constitutes the most active pole in the industry. They typically pull from clear scenario demands, find the optimal balance between performance and cost, and pursue the fastest engineering implementation and commercial closed-loop.

They often avoid head-on competition with giants on general technology platforms, instead focusing deeply on segment markets like logistics warehousing, specific manufacturing steps, and commercial services. For example, Agility Robotics' Digit robot has achieved mass production at its RoboFab factory and entered e-commerce logistics centers; Figure AI is conducting pilot tests on production lines in cooperation with manufacturers like BMW.

The rise of Chinese companies is an important phenomenon in this pole. Leveraging advantages in supply chain, cost control, and rapid iteration, Chinese companies like Unitree and Fourier are transforming from "technology followers" to "important market participants," showcasing unique competitiveness in global competition.

Table 3-3: Commercialization Progress of Representative Companies in the Scenario-Driven Camp (Early 2026)

3.4 The Key Supply Chain: The "Invisible Battlefield" Determining Mass Production Fate

Whether humanoid robots can achieve mass production of tens or even hundreds of thousands of units annually ultimately depends on the maturity of the supply chain. The modular, quickly replaceable joint design adopted by the all-electric Atlas itself foreshadows a supply chain revolution.

• Core Components: High torque-density motors, precision reducers, torque sensors, tactile sensors are key to cost and performance. Hyundai Mobis' cooperation with Boston Dynamics to develop actuators is precisely to control this core.
• Dexterous Hands: Once a cost "black hole," but companies like China's DexForce have reduced the cost of high-performance dexterous hands to one-fifth of international equivalents, significantly breaking the mass production bottleneck.
• Specialized Chips: Edge AI chips provided by NVIDIA, Texas Instruments (TI), Qualcomm, etc., are the "computing heart" of robot intelligence. Their performance, power consumption, and cost directly determine the capability boundaries of the robot body.

3.5 Landscape Summary: A New Normal of Convergence and Coopetition

In summary, the current industrial landscape exhibits distinct characteristics of "Two Poles with One Axis":

• One pole is the "Body" Expert (Boston Dynamics), relying on top dynamics control to root downwards into industrial scenarios.
• The other pole is the "Brain" Giant (NVIDIA, etc.), relying on AI and computing power to build universal platforms upwards.
• In the middle are the active "Scenario" Startups, horizontally develop the market with flexibility and speed.
• The "Axis" running through it all is the Supply Chain, which determines whether all grand visions can be realized as affordable products.

Their relationship is no longer simple competition but deep convergence and coopetition. For example, Boston Dynamics uses NVIDIA's chips and models, while startups also commonly rely on NVIDIA's ecosystem for development. Future winners are likely not the complete victory of any single pole, but those ecosystems or enterprise alliances that can most effectively integrate the reliability of "Body," the intelligence of "Brain," the pragmatism of "Scenario," and the economics of the "Supply Chain".

4.1 Track A: High-End Industrial Labor – Manufacturing's "Productivity Revolution 2.0"

This is the main battlefield with the fiercest current competition and the most direct value validation. Its core logic is to leverage the general-purpose form of humanoid robots to fill high-risk, repetitive (often referred to in the industry as "3D jobs": Dangerous, Dirty, Dull) and skilled labor shortage positions, achieving deterministic returns on single-point investments.

Application Deepening: Applications have expanded from initial demonstrative material handling to more complex production stages. For instance, in a Cummins engine factory, robots have stably taken on the task of moving material boxes. A more advanced case from collaboration between a Chinese robotics firm and an automotive parts manufacturer shows expansion from 1 workstation to covering 4 workstations across 3 different assembly lines, with material types increasing from 4 to over 20, and carrying capacity rising from 5-6 kg to 14 kg.

Efficiency Validation: According to data from a major robotics research consortium, humanoid robots can increase production line efficiency by 40%, with an investment payback period shortened to 2-3 years, providing solid financial justification for large-scale procurement.

Scale Prediction: The market is transitioning from "single-digit validation" to "tens-of-thousands mass production." Multiple institutions predict that the domestic humanoid robot industry will cross the inflection point from "1 to 10" to "10 to 100" scaling in 2026. Data from Gaogong Industry Research Institute (GGII) indicates that domestic shipments are expected to reach 62.5 thousand units in 2026, while some experts are more optimistic, predicting a level of 100 to 200 thousand units. This is backed by concrete orders, such as UBTech's humanoid robot order value approaching ~$200 million in 2025.

Table 4-1: Progress and Value Validation Cases in Industrial Applications (2026)

4.2 Track B: Commercial and Consumer Services – The Vast Frontier Towards "Ubiquitous Intelligence"

This is the track with the broadest market imagination, but also the most demanding in terms of robot cost, reliability, and human-robot interaction. In 2026, this track is accelerating its penetration from early "experiential marketing" towards creating substantive business value.

Cost Reduction and Miniaturization: Price is the key threshold for commercial success. The industry consensus is that large-scale entry into households requires prices to drop below $30,000 or even $15,000. In 2026, miniaturized, low-cost products became a new trend: Unitree Robotics' Unitree R1 starts at $4,000, and other startups have introduced robots priced under $1,500, greatly activating trial demand in the consumer-grade market.

Scenario Diversification and Deep Exploration:

1. Commercial Retail & Service: Robots like those from emerging startups have begun working in "Robot MART" retail kiosks in core business districts of major cities, not just as sales tools but also as carriers to enhance consumption experience and create emotional value. In civil service halls, robots can deeply participate in the entire process from greeting to witnessing oaths.
2. Special Services & Logistics: Robots are entering nursing homes to practice care services and accumulate the industry's first high-quality elderly care operation datasets. In the "last 100 meters" of logistics delivery, unmanned delivery vehicles like Neolix's models have achieved scaled deployment, with cumulative units exceeding 16,000 and total mileage nearing 50 million miles, solving the efficiency and cost issues of last-mile delivery.

Model Innovation: Business models are also evolving. For example, some robotics companies have launched the industry's first robot rental platform, "Qingtian Zu," lowering user barriers through leasing models and accelerating market education.

4.3 Track C: Developer Platforms & Simulation Ecosystems – The "Infrastructure" Determining Industry Height

This is the implicit track that determines the innovation speed and "ceiling" of the entire industry. Its value logic is to lower the development threshold for the entire industry by providing standardized tools and platforms, collecting "empowerment tax" or "ecosystem fees."

• Core Value: As demonstrated by NVIDIA at CES 2026, its goal is to build an "Android" ecosystem for robot development through the GR00T model, Cosmos World Model, and Omniverse simulation platform. This addresses two long-standing core pain points in robotics development: data scarcity and the simulation-to-reality gap. Reports from the International Federation of Robotics (IFR) also note that generative AI and world models can generate training data through simulation, enabling robots to learn new tasks autonomously, which is key to moving towards autonomy.
• Ecosystem Effect: The platformization strategy will attract global developers, research institutions, and中小 startups to develop applications and innovate based on unified standards. This not only accelerates technology diffusion but also builds a strong moat for the platform builder. As pointed out by industry analysts, the core differentiator of future robot competitiveness will entirely depend on the maturity of software ("Brain" & "Cerebellum"), and the platform is the cornerstone accelerating this process.

Table 4-2: Comparison of Key Characteristics and Driving Logic of the Three Major Business Tracks (2026)

Conclusion: The Concerto and Division of a Trio

The three tracks are not distinctly separated but mutually reinforcing and co-evolving. Industrial applications (Track A) provide the most solid cash flow and scenario refinement opportunities, fueling technological progress. Commercial services (Track B) depict the ultimate broad market, pulling technology towards low-cost, high-intelligence iteration. Platform ecosystems (Track C) provide the accelerator for innovation in the first two, determining the overall evolution speed of the industry.

The all-electric Atlas chose to first delve deeply into Track A, the optimal solution combining its technological DNA with shareholder (Hyundai Motor Group) resources. In contrast, Tesla Optimus and Figure AI attempt to find balance between Tracks A and B. In the future, companies that can establish absolute advantage in one track and successfully radiate their capabilities to another track are most likely to become the ultimate winners in this value creation race.

5.1 Technical Convergence: From "Brain and Body" to "Autonomous Collaborator"

Future robots will no longer be a simple assembly of an AI "brain" and a dynamics "body." Instead, under the framework of "Physical AI," they will evolve into intelligent agents with autonomous decision-making and collaboration capabilities through layered and synergistic technology.

1. Layered Collaborative Architecture Between AI and Dynamics Becomes Standard
A consensus technical architecture is emerging in the industry:

• Decision Layer (AI Brain): Driven by multimodal world models and embodied intelligence large models. They are responsible for high-level task understanding, semantic reasoning, and rough planning. For example, Boston Dynamics' Atlas has integrated Google's Gemini Robotics large model, upgrading from "command execution" to "autonomous decision-making."
• Execution Layer (Control Cerebellum): Based on classical dynamics and real-time control technology, ensuring action precision, stability, and safety. As demonstrated by the new Atlas, top motion control capabilities are being packaged into reliable, reusable modules.
• Perception & Interaction Layer: Fuses multimodal information like vision, language, and touch, allowing robots to understand and adapt to the physical world more delicately.

Table 5-1: Key Technological Manifestations of AI-Robotic Dynamics Convergence in 2026

2. "Vehicle-Robot-Chip" Synergistic Ecosystem Accelerates Industry Maturation
The automotive and robotics industries share high commonality in **high-performance chips, high-reliability requirements, and complex system integration**. In 2026, **"automotive-grade technology migrating to robotics"** has become a key path for cost reduction, efficiency improvement, and accelerated mass production.

• Chip Reuse: NVIDIA DRIVE Thor (automotive) and Jetson Thor (robotics) share architectures; Qualcomm repurposes its Digital Chassis technology for robotics platforms; domestic chipmakers are transferring automotive-grade chip experience to robotics collaborations.
• Manufacturing & Supply Chain Synergy: The Hyundai Motor Group providing mass production capability for Boston Dynamics' Atlas is, in essence, the empowerment of the robotics industry by automotive manufacturing systems.

5.2 Industry Screening: From "A Hundred Contenders" to the "Mass Production Delivery" Clearance

While technological convergence provides the engine for industrial development, the market itself is initiating a cruel but necessary screening. In 2026, the industry will accelerate from a "hundred contenders war" into a "clearance" phase where mass production and delivery capability serve as the litmus test.

• Current State: Crowded Tracks, Cautious Capital. According to reports from institutions like the Beijing Academy of Artificial Intelligence (BAAI), there are currently over 200 embodied AI companies in China alone, with more than 100 focused on humanoid robots. However, the technical difficulty and funding requirements of this field far exceed the internet's "group-buying wars," while the capital environment is more severe. The number of companies already exceeds the circuit's physical capacity and capital supply.
• Turning Point: Clients Shift from "Universities" to "Factories". A decisive change is occurring: from 2025 to 2026, the core customer base is rapidly shifting from universities and research institutions to B-end industrial clients with real needs. This means the metric changes from "papers and demos" to "orders and ROI."
• Key to Winning: Closed-Loop Evolution & Commercial Landing Capability. Those that can survive and win in this reshuffle will be companies with scenario closed-loop capability. This includes: 1) Engineering capability to translate technology into stable, reliable products; 2) Replicable business model in specific scenarios (e.g., auto manufacturing, logistics); 3) Evolutionary capability for continuous product iteration through real data feedback.

Table 5-2: Evolution of Humanoid Robot Industry Stages and Key Characteristics in 2026

5.3 Ultimate Challenges: The Final Barriers Before Scaling

Even with clear trends, the industry still faces several high walls that must be collectively scaled on its path to true scale.

1. Technical Bottlenecks: Generalization Capability from "Usable" to "Easy to Use"
   • Scenario Generalization: Current robots perform well on specific, debug production lines, but their adaptability to unstructured environments and untrained long-tail tasks remains insufficient.
   • Cost-Reliability Balance: Ensuring industrial-grade reliability for tens of thousands of hours of failure-free operation while reducing the Bill of Materials (BOM) cost below ~$15,000 is a massive engineering challenge.
2. Safety & Ethics: The Invisible "Guardrails" Must Be Fortified
   • Functional Safety: As human-robot collaboration deepens, robots must pass strict international safety certifications like those from ISO.
   • Data & AI Safety: The vast amounts of environmental and interaction data collected by robots involve privacy; the explainability and responsibility of AI "black box" decisions in safety-critical scenarios urgently require legal and ethical frameworks.
3. Social Acceptance: Narrative Shift from "Replacer" to "Collaborator"
   • While filling global labor gaps, robots, especially humanoids, will inevitably raise concerns about job displacement. Successful promotion requires collaboration between businesses, governments, and workers to position robots as **assistants** taking on "dangerous, dirty, dull" tasks, and helping the workforce transition through skills retraining.

6.1 Core Conclusions: Paradigm Shift Complete, The Era of Value Creation Begins

The arguments of this report ultimately point to three irreversible core conclusions:

1. Technologically, the deep integration of "Body" and "Brain" is a foregone conclusion.
The historical mission of model-based dynamics and control, represented by Boston Dynamics, has shifted from "front-stage spectacle" to "backstage foundation." It has not been eliminated but has been engineered and modularized, becoming the "autopilot" ensuring safe, reliable, and precise robot action execution. Meanwhile, AI large models and Physical AI, propelled by companies like NVIDIA, Google, and Tesla, are taking on the roles of the "cognitive brain" responsible for high-level task understanding, planning, and generalized decision-making. The two are tightly coupled through a layered architecture (decision-planning-execution), each indispensable. The future competitive focus will be on who can achieve this integration more elegantly and efficiently.

2. Industrially, systemic capability has replaced single-point technology as the key determinant of success.
The competitive dimensions of the industry have escalated from "who can do a backflip" a decade ago to a comprehensive contest encompassing core technical depth, engineering and mass production capability, AI algorithm level, scenario landing closed-loop, supply chain cost control, and ecosystem building capability. Leading in a single dimension is no longer sufficient to build a long-term moat. For instance, Boston Dynamics has filled its mass production and scenario gaps with the support of the Hyundai Motor Group; emerging startups must quickly prove that their technology can translate into deliverable products at scale and calculable commercial returns.

3. Commercially, pragmatism centered on ROI is now fully dominant.
The patience of capital markets and end customers is waning. In 2026, the industry's evaluation criteria have completely shifted from "amazing demo videos" to "clear customer orders," "verifiable investment payback periods (typically requiring 2-3 years)," and "replicable, scalable application scenarios." Technology must answer the fundamental question: "Who does it solve problems for, what problems, and how much value does it bring?" Industrial scenarios have become the main battlefield precisely because their value calculation is the most direct.

6.2 Investment and Strategic Framework

Based on the new industry norm of "technological convergence, systemic competition, and pragmatic commerce," we construct the following strategic framework for different participants:

Table 6-1: Robotics Industry Participant Investment and Strategic Framework (2026 Perspective)

6.3 Risk Warnings and Future Outlook

While embracing opportunities, it is essential to recognize that the industry is still in the early stages of scaling and faces multiple risks:

1. Technology Iteration Risk: AI and robotics hardware technologies are still evolving rapidly. Currently leading architectures or solutions may be disrupted by completely different technological paths within a few years.
2. Commercialization Underperformance Risk: Current optimistic shipment forecasts are based on early pilots. Large-scale deployment may reveal unforeseen challenges in cost, reliability, maintenance, etc., leading to slower market penetration growth.
3. Geopolitical & Supply Chain Risk: The globalized layout of core chip and precision component supply chains faces uncertainty, potentially affecting the stable development and cost control of the global industry.
4. Ethical & Regulatory Risk: As robots integrate deeply into society, issues like data security, privacy protection, and human-robot liability will trigger increasingly strict regulation, increasing compliance costs.

Future Outlook (2026-2030)

We anticipate the industry will follow a path of "first convergence, then diffusion":

• Near Term (1-2 years): Industry accelerates consolidation, resources concentrate towards leading companies and players with real scenarios. The first star products with annual deliveries exceeding ten thousand units and dominant development platforms will emerge.
• Medium Term (3-5 years): After a stable industrial landscape emerges, technology costs decrease further, and applications begin to spread to broader fields such as commercial services and specialized operations, potentially creating new platform opportunities..
• Long Term: Robots deeply integrate with ambient intelligence and the Internet of Things, ultimately moving towards general embodied intelligence. However, its realization path and timeline remain highly uncertain.