Humanoid robotics enters the 2026 technology frontier: World Action Models may reshape the competitive landscape for robot brains
AI summary card
Humanoid robotics enters the 2026 technology frontier: World Action Models may reshape the competitive landscape for robot brains
Bernstein believes humanoid robotics is moving from demonstrations toward the eve of large-scale commercialization, with the technical focus shifting from stable walking to full-body high-dynamic control, long-horizon manipulation, autonomous planning, and world models represented by WAM.
- The technology frontier is summarized across five dimensions: motion, manipulation, autonomy, robot brain models, and data modalities.
- VLA is evolving toward World Action Models, where WAM generates actions by predicting physically feasible future states, potentially improving task generalization and robustness.
- Competitive moats may shift partly away from simply owning teleoperation or deployment data toward model architecture, data integration capability, and the ability to fill critical physical-data gaps.
- The report reiterates Outperform on FANUC, Inovance, Harmonic Drive, Cognex, and Keyence, while Estun remains Market-Perform.
Report interpretation
Overview
This report focuses on the humanoid robotics theme within global automation, outlining the industry's technological frontier and future direction in 2026. The report argues that humanoid robots have advanced from the impressive flat-ground stable walking of two years ago to a stage requiring full-body high-dynamic control, environmental adaptation, complex manipulation, and higher autonomy. Although long-horizon tasks, industrial-grade reliability, and general autonomous planning are still at an early stage, the industry is entering an accelerated development phase and approaching the threshold for large-scale commercial adoption.
Core views
The core view is that the next phase of competition in humanoid robotics will center on motion capability, manipulation capability, autonomous planning, robot brain models, and data modalities. Robot brains are evolving from LLM/VLM to VLA, and then toward post-2026 WAM. WAM does not merely map language to action; instead, it predicts physically feasible future states based on environmental state, context, and experience, and then generates actions. This paradigm may reduce dependence on expensive teleoperation data and hardware-specific deployment data, making model architecture, cross-platform data integration, and non-visual data capabilities such as touch into new moats.
Analysis framework
The report uses a combination of technology roadmap analysis and case benchmarking: it first defines the humanoid robotics frontier across five dimensions, then lists the latest demonstrations or model progress from companies such as Unitree, Agibot, LimX, Paxini, Figure AI, Boston Dynamics, Physical Intelligence, and Nvidia, and explains technical differences through comparisons of VLA, world models, diffusion policy, and WAM. On the investment side, it evaluates robot-brain suppliers, complete-unit OEMs, sensor/data gap fillers, and listed industrial automation companies within the same industry chain framework.
Methodology notes
Evolution of robot brain models
VLA can be understood as 'language to action,' while WAM places more emphasis on predicting future states based on environmental state, context, and experience, and then generating actions; the report believes WAM could improve complex-task generalization, skill transfer, and performance robustness.
Three categories of world model applications
The report divides world models into three categories: renderer, simulator, and planner. Among these, the planner is closer to the capabilities needed for humanoid robot action decisions because it must produce physically feasible future states and action plans.
Primary valuation method for covered companies
The disclosure section shows that target prices for Estun, FANUC, Harmonic Drive Systems, Keyence, Inovance, and Cognex mainly use EV/EBITDA multiples, with DCF as a reference for long-term intrinsic value.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- FANUC Corp (6954.JP)A listed company tied to industrial automation and robotics; the report reiterates Outperform.
- Strengths
- Beneficiary of the global automation and robotics capex cycle, with a target price of JPY7,000.
- Weaknesses
- End markets are cyclical and affected by the global macroeconomy and exchange rates.
- Comparison
- Mapped as a secondary-market proxy within the robotics value chain alongside other Japanese automation companies.
- Risks
- Global automation demand weaker than expected, market share loss, and JPY appreciation.
- Shenzhen Inovance Technolo-A (300124.CH)A core Chinese automation name; the report reiterates Outperform.
- Strengths
- Has a market-share gain thesis in China's industrial automation, servo motors, and VFDs, with a target price of RMB82.00.
- Weaknesses
- Sensitive to China's automation demand, industrial capex, and EV demand.
- Comparison
- Rated more positively than Estun; both are proxies within the Chinese automation chain.
- Risks
- China automation demand weaker than expected, slower-than-expected share gains in non-servo and VFD segments, and weaker-than-expected EV demand.
- Harmonic Drive Systems Inc (6324.JP)A key supplier of robot reducers; the report reiterates Outperform.
- Strengths
- Holds over 50% global market share in strain wave reducers, with a target price of JPY7,800.
- Weaknesses
- Changes in the competitive landscape are especially critical to it.
- Comparison
- Compared with general automation companies, its risk is more concentrated in robot demand and reducer competition.
- Risks
- Global robot demand weaker than expected, competitors taking share, and JPY appreciation.
- Keyence Corp (6861.JP)A global automation sensing and control-related name; the report reiterates Outperform.
- Strengths
- Linked to manufacturing capacity utilization in major economies, with a target price of JPY86,000.
- Weaknesses
- Near-term growth volatility may be affected by changes in capacity utilization.
- Comparison
- Similar to Cognex, with risks mainly tied to global manufacturing utilization and automation demand.
- Risks
- Capacity utilization weaker than expected, delayed or below-expected global automation demand, and JPY appreciation.
- Cognex Corp (CGNX)A machine vision and automation-related name; the report reiterates Outperform.
- Strengths
- Target price of USD75.00, benefiting from a recovery in global automation demand and progress in new customer projects.
- Weaknesses
- Sensitive to the global macroeconomy and manufacturing capacity utilization.
- Comparison
- Similar to Keyence, both mapping automation sensing and machine vision demand.
- Risks
- Global automation demand weaker than expected, slower-than-expected progress in emerging customer projects, and slower-than-expected Moritex integration and synergy realization.
- Estun Automation Co Ltd (002747.CH / 2715.HK)A China industrial automation and robotics-related name, rated Market-Perform by the report.
- Strengths
- There is upside potential from margin expansion and market share gains.
- Weaknesses
- High pressure on valuation and operational improvement delivery, with target prices below the current prices shown in the table.
- Comparison
- Within the report's covered group, it is rated lower than FANUC, Inovance, HDSI, Keyence, and Cognex.
- Risks
- China automation demand weaker than expected, delayed or slower-than-expected improvement in operations and net margin, slower-than-expected market share gains, and weaker-than-expected Cloos integration and synergies.
- NvidiaA potential beneficiary in the independent robot brain and foundational WAM ecosystem.
- Strengths
- The report mentions the DreamZero model and believes the WAM paradigm may favor third-party suppliers with model architecture and data integration capabilities.
- Weaknesses
- This company is covered by Bernstein's U.S. semiconductor team, and the current report does not provide a detailed rating or target price within this piece.
- Comparison
- Similar to Physical Intelligence, it is more of a robot-brain or foundation-model supplier than an integrated brain-body OEM.
- Risks
- WAM inference speed, scarcity of non-visual data, and the pace of industry adoption may affect commercialization.
Key data
- Report date2026-07-20Both the file name and metadata date point to July 20, 2026.
- Core technology dimensionsMotion, manipulation, autonomy, robot brain models, data modalitiesThe report says the humanoid robotics frontier and future direction can be evaluated along these five dimensions.
- Ratings on major listed namesFANUC, Inovance, Harmonic Drive, Cognex, and Keyence are rated Outperform; Estun is rated Market-PerformThe report explicitly reiterates these ratings in the investment implications section.
- Target pricesEstun A RMB26.00；Estun H HKD17.26；FANUC JPY7,000；HDSI JPY7,800；Keyence JPY86,000；Inovance RMB82.00；Cognex USD75.00From the Bernstein ticker table and valuation methodology disclosure.
- Current challenges for WAMInference is slower than VLA, and non-visual training data such as tactile and material properties are extremely limitedThe report identifies these as the main shortcomings for real-world deployment of World Action Models.
Impact & implications
The investment implication is that the competitive focus in the humanoid robotics value chain may expand from complete units and data ownership to third-party robot brains, foundational WAM, cross-platform data integration, and physical-data gap filling. Independent robot-brain or foundation-model developers such as Physical Intelligence and Nvidia may benefit and complement integrated brain-body OEMs such as Figure AI, Agibot, and LimX. Tactile-sensing companies such as Paxini may also gain strategic value by filling physical-data gaps in world models. For the secondary market, the report continues to favor automation- and robotics-related names such as FANUC, Inovance, Harmonic Drive, Cognex, and Keyence, though macro automation demand and valuation cycles still warrant monitoring.
Risks
- A weaker global and Chinese macroeconomy could suppress industrial automation capex.
- Global robot demand or automation demand could come in below expectations.
- Trade frictions, exchange-rate volatility, and JPY appreciation could affect related Japanese names.
- Humanoid robot long-horizon manipulation, industrial-grade reliability, and autonomous planning remain at an early stage, creating uncertainty around the pace of commercialization.
- Compared with VLA, WAM has slower inference speed and insufficient non-visual data such as tactile and material-property data.
- Key companies may face market share losses, weaker-than-expected M&A integration, or slower-than-expected synergy realization.
What to watch
- Whether WAM can continue improving generalization, skill transfer, and robustness in real complex tasks.
- Evidence of reliability as humanoid robots move from short demonstrations to continuous operation in industrial scenarios.
- The scaled availability of non-teleoperation, platform-agnostic data, especially first-person human video and tactile data.
- Cooperation models between third-party robot brain or data gap-fillers such as Physical Intelligence, Nvidia, and Paxini and complete-unit OEMs.
- Changes in automation demand, orders, margins, and market share for FANUC, Inovance, HDSI, Keyence, Cognex, and Estun.
- Whether the EV/EBITDA multiples implied by target prices are re-rated along with cycles and competitive trends.