Humanoid robots approach the eve of commercialization, with WAM becoming the core variable of the next-generation robot brain

Overview

This report focuses on the technological frontier of humanoid robots within global automation. Bernstein judges that humanoid robots are entering an accelerated development phase and are near the threshold of large-scale commercial adoption. The report characterizes the frontier through five dimensions: motion capability, manipulation capability, autonomy, robot-brain models, and data modalities, and uses cases such as Unitree, Agibot, LimX, Figure AI, Boston Dynamics, Physical Intelligence, and Nvidia to illustrate industry progress.

Core views

The report’s core view is that the key change in humanoid robots is not just improved hardware motion capability, but the evolution of the robot brain from VLA to WAM. VLA is more like “language to action,” whereas WAM predicts physically feasible future states based on environmental state, context, and experience, and then generates future actions. This paradigm is expected to improve task generalization, skill transfer, and performance robustness. At the same time, WAM emphasizes the diversity of action and environmental data, which may make open-source, cross-platform, human-centric video data and non-visual data more important, thereby altering value-chain moats.

Analysis framework

The report first defines the technological frontier of humanoid robots through a five-dimensional framework, then compares industry progress using representative companies and model cases; it then explains how technical approaches such as world models, WAM, and diffusion policies affect robot autonomous planning and task generalization; finally, it turns to implications for automation stock investing, valuation methodology, and company-level risks.

Methodology notes

• Technical frameworkFive-dimensional frontier framework for humanoid robots

  Motion, manipulation, autonomy, robot brain, data modalities

  The report uses five dimensions to measure the progress of humanoid robots as they move from early demonstrations toward commercial capabilities, including highly dynamic whole-body motion, long-horizon dexterous manipulation, short-term and long-term autonomous tasks, evolution of robot-brain models, and a shift from teleoperation data toward more cross-platform data.

• Model paradigmWorld Action Model

  WAM as the planner of the robot brain

  WAM generates subsequent actions by predicting physically feasible future environmental states, similar to how the human brain imagines outcomes before acting; the report believes this helps task generalization, skill transfer, and robustness, but it still faces issues with inference speed and insufficient non-visual data.

• Model comparisonComparison of VLA and WAM

  From “language to action” to “action based on future states”

  VLA directly outputs the next action, while WAM first predicts how the environment will evolve and then plans actions accordingly; this makes WAM better suited to complex, multi-solution, and long-horizon tasks.

• Valuation methodologyEV/EBITDA and DCF reference

  Target-price setting for automation stocks

  The disclosure section shows that covered companies mainly use EV/EBITDA multiples as the target-price methodology, with DCF as a reference for long-term intrinsic value; target multiples are adjusted based on historical cycles, long-term trends, and competitive dynamics.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Independent robot-brain or foundational WAM developers such as Physical Intelligence and Nvidia

  May benefit from the shift to the WAM paradigm

  Strengths

  WAM requires model architecture, data integration, and cross-platform generalization capabilities, allowing independent model developers to potentially serve multiple robot-body platforms more easily.

  Weaknesses

  They need to prove that inference speed, task stability, and non-visual data capabilities can meet real deployment requirements.

  Comparison

  Relative to integrated brain-body OEMs, their advantage may be closer to that of general-purpose technology platforms and third-party suppliers.

  Risks

  If OEMs more quickly establish an in-house closed loop, the bargaining power and penetration of third-party robot-brain suppliers may be constrained.

• Integrated brain-body robot OEMs such as Figure AI, Agibot, and LimX

  Core participants in commercialization and data closed loops

  Strengths

  They can optimize the robot body, actuators, sensors, task scenarios, and robot brain in an integrated manner.

  Weaknesses

  They may need to continuously invest in expensive hardware, deployment, and data-collection systems.

  Comparison

  Relative to third-party WAM developers, OEMs are closer to applications and real-world scenarios, but scaling general-purpose models may be harder.

  Risks

  There is uncertainty around mass-production reliability, cost reduction, long-horizon autonomous task success rates, and demand from commercial scenarios.

• Companies related to tactile and physical data such as PaXini

  May fill a critical physical-data gap in WAM training

  Strengths

  Non-visual data such as touch and material properties are important supplements for world models to understand physical environments.

  Weaknesses

  The current scale of non-visual data is limited, and ecosystem maturity remains low.

  Comparison

  Compared with pure visual or video data, tactile data is scarcer and more likely to form specialized value.

  Risks

  If visual data and simulation data are sufficient to support most tasks, the strategic importance of tactile data suppliers may be lower than expected.

• FANUC、Inovance、Harmonic Drive、Cognex、Keyence

  Related stocks in the automation value chain with reiterated Outperform ratings

  Strengths

  They cover industrial automation, motion control, precision reducers, machine vision, and sensing automation, and may benefit from long-term automation and robotics trends.

  Weaknesses

  Different companies have varying degrees of direct revenue exposure to humanoid robots, and near-term earnings are still affected by traditional automation cycles.

  Comparison

  The report rates all of these names higher than Estun.

  Risks

  Global automation demand weaker than expected, intensifying competition, yen appreciation, weak China demand, or a downturn in the industry capex cycle.

• Estun

  Automation-related name with reiterated Market-Perform rating

  Strengths

  It has exposure to China’s automation and robotics value chain.

  Weaknesses

  The report’s target price is below the current price disclosed in the table, and its rating is lower than other major automation names.

  Comparison

  Relative to the Outperform names, the report’s investment view on Estun is more neutral.

  Risks

  China automation demand weaker than expected, weaker-than-expected Cloos integration and synergies, or slower-than-expected margin improvement or market-share gains.

Key data

• Report date2026-07-20The filename date is 20260720, and the table price date is 16 Jul 2026.
• AuthorsJay Huang, Ph.D.；Weibin Liang, Ph.D.；Dien Wang, Ph.D.The report cover discloses three authors and their contact information.
• Frontier technology casesUnitree、Agibot、LimX、PaXini、Figure AI、Boston Dynamics、Physical Intelligence、NvidiaThe report uses representative companies to demonstrate progress in highly dynamic motion, environmental adaptation, long-horizon manipulation, reliability, and robot-brain models.
• Key modelsUnifoLM-WMA-0、COSA 0.5、Helix 02、π0.7、DreamZeroThese models or platforms are used to illustrate progress in humanoid robot brains and world-model directions.
• Stock ratingsFANUC、Inovance、Harmonic Drive、Cognex、Keyence 为 Outperform；Estun 为 Market-PerformThe report explicitly reiterates the ratings in the investment implications section.
• Target pricesEstun A-shares RMB26.00，Estun H-shares HKD17.26，FANUC JPY7,000，Harmonic Drive JPY7,800，Keyence JPY86,000，Inovance RMB82，Cognex USD75.00Source: the disclosure section and Bernstein ticker table.

Impact & implications

If WAM becomes the mainstream paradigm, competition in humanoid robots may shift away from pure reliance on proprietary teleoperation data and embodied deployment data toward model-architecture capability, cross-platform data integration capability, and the ability to fill gaps in physical-world data. Independent robot-brain or foundational WAM developers may become key links similar to electric-vehicle battery and powertrain suppliers; integrated brain-body OEMs may still develop core systems in-house, but some OEMs may procure the best third-party robot-brain capabilities.

Risks

• WAM currently has slower inference speed than VLA, which may limit real-time deployment in complex tasks.
• Non-visual data such as touch and material properties is very limited, which may constrain world models’ understanding of real physical environments.
• Humanoid robots’ manipulation capability, long-horizon autonomy, and industrial-grade reliability are still at an early stage, and commercialization may proceed more slowly than optimistic expectations.
• Global industrial automation demand is cyclical, and macroeconomics, capex cycles, and manufacturing capacity utilization will affect the performance of related stocks.
• Changes in the competitive landscape may affect the market share and pricing power of key component companies such as Harmonic Drive.
• Exchange rates, trade friction, M&A integration, and regional demand fluctuations are common risks across multiple covered companies.

What to watch

• Whether WAM inference speed can approach the requirements of real robot deployment.
• The success rate of long-horizon autonomous planning, skill transfer, and task generalization in real industrial or household scenarios.
• Whether humanoid robots move from short-term demonstrations to continuous operation, industrial-grade reliability, and replicable commercial scenarios.
• Whether robot training data accelerates its shift from teleoperation data toward egocentric video, platform-agnostic data, and non-visual physical data.
• Whether value-chain specialization emerges between independent robot-brain suppliers and integrated brain-body OEMs.
• Whether the target prices, ratings, and automation-demand assumptions for FANUC, Inovance, Harmonic Drive, Cognex, Keyence, and Estun change.