The inflection point of "physical AI" at CES: Robotaxis move towards large-scale deployment, and the supply chain for humanoid robots quietly takes shape

Deutsche Bank predicts that this year will be the first year for the large-scale deployment of Robotaxis and humanoid robots. Humanoid robots are leveraging the automotive supply chain to accelerate cost reduction, with manufacturers like Mobileye setting a cost target of $10,000. Autonomous driving is moving from the testing phase to large-scale implementation. Waymo has seen a surge in orders, and NVIDIA has launched the Alpamayo platform to help automakers achieve "plug-and-play" capabilities. Physical AI is moving from the laboratory to mass production.

2026 might mark the beginning of AI's large-scale entry into the physical world - from walking robots to autonomous vehicles, AI is accumulating ecological hardware.

According to information from Zhui Feng Trading Desk, a research report released by Deutsche Bank on January 13th showed that the bank's analyst team attended the CES exhibition in Las Vegas last week and felt a significant surge in market enthusiasm and relevance. The bank pointed out that vehicle autonomous driving (Robotaxi + consumer-grade L4) and, most notably, humanoid robots took center stage at the exhibition.

Deutsche Bank summarized in the report: "Overall, we predict that 2026 will be a year when autonomous vehicles increasingly transition from testing/validation to large-scale implementation, while humanoid robots will move from laboratory experiments to small-scale deployment."

The report emphasized that a new supply chain is being cultivated in the field of humanoid robots, and suppliers are trying to transform into this field in anticipation of large-scale production in the future. Meanwhile, the deployment of Robotaxis in the autonomous driving field is gaining strong momentum, and chip giants like NVIDIA are reshaping the competitive landscape by launching new platforms.

Deutsche Bank listed 10 core observations in the report:

01

The supply chain for humanoid robots is taking shape

Actuators are becoming the "muscle" entry point

Deutsche Bank believes that although it is still in the early stage, suppliers have started to shift towards the humanoid robot supply chain, following a path similar to that of electric drive systems: providing both integrated solutions and underlying components.

• Schaeffler is attempting to become the main "muscle" for humanoid robots, offering linear and rotary actuators.

At CES, it showcased an integrated planetary gear actuator for humanoid robots: a compact integration of a two-stage planetary gearbox, motor, encoder, and controller. This unit features high thermal stability, a torque range of 60–250 Nm, and low back-driving ability, which can withstand external forces and prevent accidental reverse rotation of the drive components, making it suitable for continuous operation. Deutsche Bank mentioned that NEURA has agreed to use Schaeffler actuators in its humanoid robots, and it seems that other customers have already used (at least some components) or will use them in the future.

• Hyundai Mobis also announced that it will supply actuators for Boston Dynamics' Atlas

, aiming to enable robots to be manufactured using the automotive large-scale supply chain.

When the supply chain starts to "automotivize", it is often not the concept but the penetration of key components and large-scale manufacturing capabilities that are priced first.

02

The landscape of on-board chips

NVIDIA remains the top choice, but differentiation is starting to emerge

Deutsche Bank observed that NVIDIA still dominates the on-board processors for humanoid robots, mainly due to performance and ease of use. Companies using Jetson Orin or Thor include: 1X, Agility, Apptronik, Boston Dynamics, Figure AI, Mentee, (currently) NEURA, UBTECH, Unitree, etc.

In contrast:

• Tesla and XPeng use self-developed inference chips.
• At CES, Qualcomm launched a next-generation solution for the "full-stack architecture" of robots

(Dragonwing IQ10 Series), but Deutsche Bank said it is unclear whether it will be widely adopted by customers; meanwhile, VinMotion's Motion 2 humanoid robot uses the IQ9 Series, and the IQ10 was initially designed for industrial AMRs and more advanced full-size humanoid robots.

03

"Physical AI" is moving from scripts to agents

VLA is becoming the main line

One of the most significant paradigm shifts on-site is the move from "pre-programmed/scripted actions" to vision-language-action (VLA), enabling robots to "reason" and complete tasks.

Boston Dynamics replaced the traditional MPC (model predictive control) with Google DeepMind Gemini Robotics' VLA model, enabling Atlas to understand previously unseen environments (such as chaotic scenes in unstructured factories).

Its action execution is supplemented by TRI's large behavior model (LBM), similar to Figure's Helix dual-system model: System 1 provides high-frequency and rapid responses, while System 2 conducts high-level reasoning and language processing at low frequencies; Deutsche Bank also pointed out that Figure seems to be developing two sets of models on its own.

04

The training competition is intensifying

Real-world data

The "closed-loop" with simulation is the key

Deutsche Bank believes that the industry debate has shifted from "which is better, simulation or the real world" to "how to achieve an efficient closed-loop".

• NEURA is taking a more "physics-first" approach

, building the NEURA Gym, a large-scale physical training center. It believes that simulation is an "approximation" and will be inaccurate in complex contact scenarios (such as "threading a needle"). It collects high-fidelity data through hundreds of robots performing real tasks such as sorting and assembly, then inputs it into "Neuraverse" to generate "synthetic twins" of real failures for training in simulation, and finally pushes the repair solutions back to real robots.

Another company mentioned that it is unable to simulate the "tactile sensation" of objects and requires human demonstration first: through remote operation, a person wears a VR suit to control a humanoid robot to perform actions such as "picking up grapes". After using a small number of "perfect examples", NVIDIA's GROOT-Mimic is used to generate "100,000+" action variants in simulation, and reinforcement learning is used to make the actions smoother.

In contrast, Mobileye emphasizes that its Mentee will be mainly trained through simulation.

05

The "general-purpose" concept gives way to "specific jobs" first

Commercialization proof takes precedence

Deutsche Bank believes that in the short term, "general-purpose humanoid robots" will be more likely to be introduced into specific scenarios to prove their commercial viability before considering entering households.

• Keenon Robotics (China)

: It already holds a 40% global market share in service robots, with a cumulative overseas export of approximately 100,000 units. Its product prices range from less than $10,000 to about $100,000, focusing on strong task customization. The highlight of CES 2026 was its flagship humanoid robot, XMAN - R1, which can make popcorn, pour drinks, and interact with anthropomorphic gestures. Its "Brain" is the Keenon Operator Model 2.0, a VLA model designed for the service industry that can understand instructions such as "find the guest at Table 4 and give him candy". Keenon also mentioned building a collaborative ecosystem at the Shanghai Shangri-La Hotel: the MAN - R1 serves as the "face" for human - robot interaction, the W3 delivers items to rooms, the S100 carries heavy luggage, and the C40/C55 performs cleaning tasks. In high - labor - cost markets like Japan, its robots have a service life of 8 years, significantly longer than the common 3 - 5 years in the industry.

• Deep Robotics focuses on industrial inspection

: Measured by the coverage distance (up to 63 km), it can perform 24/7 autonomous patrol and monitoring in dangerous areas such as substations, power plants, and oil and gas facilities. In emergency scenarios, it is used for disaster relief, fire - fighting, and toxic gas detection, and it uses replaceable batteries to reduce charging friction.

06

The cost - reduction formula is straightforward

Scale is the prerequisite for cost reduction

On the humanoid robot side, Deutsche Bank attributes the main drivers of cost reduction to: increased volume to spread costs + improved bargaining power with suppliers.

Some companies claim that the cost has dropped from "$200,000 to $100,000" and plan to reduce it to "$50,000" in the "next few years", provided that the sales volume reaches several thousand units.

Boston Dynamics and Hyundai Motor announced a target of achieving an annual production capacity of 30,000 units in 2028; and all of its 2026 production has been pre - allocated to Hyundai's automotive factories. The company also pointed out that actuators account for approximately 60% of the Bill of Materials (BoM), and this part will be manufactured by Hyundai Mobis, a supplier within the Hyundai system, to accelerate large - scale production.

Against the backdrop of Mobileye's acquisition of Mentee, it was disclosed that if the annual production volume is 50,000 units, the manufacturing cost of a relatively simplified design (without a tendon drive system) is approximately "$20,000 per unit"; if the annual production volume is "100,000 units", the cost can be halved to "$10,000 per unit", with a target of ramping up production in 2028, and production will be handled by Aumovio.

07

The momentum of Robotaxis is building up

2026 is more like an "acceleration year for commercialization"

Deutsche Bank believes that with Tesla's launch of Robotaxis in 2025, the commercialization momentum of multiple players will be stronger in 2026. The large presence of Waymo and Zoox at CES is a signal:

• Waymo:

It has provided over 10 million paid rides since its establishment. The latest disclosure shows that it reached 450,000 paid rides per week in December 2025 and expanded to Houston, Miami, and international markets such as Tokyo and London.

• Amazon's Zoox:

It has moved from public testing in Las Vegas to the display of a "market - ready product", focusing on a "carriage - style" Robotaxi for dense cities, completely without a traditional cockpit.

• Mobileye and Volkswagen

: They will launch an L4 - level Robotaxi service in Los Angeles this year using specially prepared ID. Buzz electric vans.

In addition, an autonomous vehicle project based on the Lucid Gravity, jointly promoted by partners Nuro, Lucid, and Uber, is scheduled to start in the San Francisco Bay Area at the end of 2026 and then expand to more cities.

08

20% NVIDIA Alpamayo

It packages the "brain + skull" for automakers, but validation is still underway

NVIDIA announced the launch of Alpamayo ("the brain") for autonomous driving, paired with Thor ("the skull"), aiming to lower the threshold for automakers to deploy high - level capabilities: companies like Lucid and Mercedes don't need to invest billions of dollars from scratch to build AI infrastructure and can directly "plug in" NVIDIA's solution.

Deutsche Bank remains cautious: this has indeed sparked discussions about Tesla's moat, but it's too early to be worried; NVIDIA still needs traditional OEMs to fulfill their promises, and it remains to be seen whether its model can cover real - world boundary cases. Deutsche Bank pointed out that its training data volume is only a fraction of the data collected by Tesla.

Even if Alpamayo performs ideally, Deutsche Bank still believes that Tesla has a structural cost advantage due to its vertical integration (including the entire vehicle, chips, AI infrastructure, network, etc.); if autonomous driving/Robotaxis become commoditized, cost will become the biggest differentiator.

09

Aptiv: End - to - end AI ADAS + Connectivity and Software Platform

It's about "cross - industry"

The core of Aptiv's display is the next - generation end - to - end (E2E) AI - driven ADAS platform: using the newly released Gen 8 radar and PULSE sensors to achieve L2++ hands - free driving with "human - like logic" in complex urban environments.

On the software side, it launched the cloud - native middleware platform LINC, jointly built with Wind River, to achieve a truly software - defined vehicle through 5G and C - V2X; and demonstrated with Verizon that vehicles can "see pedestrians/cyclists around the corner" by sharing real - time data. Aptiv also emphasized the expansion of sensors to the aerospace and collaborative robot fields - Deutsche Bank believes that this is the narrative that the "new Aptiv" needs to prove to strive for a re - evaluation of its valuation multiple.

10

Visteon: 700 TOPS Domain Controller, Plug - in Upgrade

It focuses on "execution"

Visteon launched the SmartCore HPC domain controller at CES, with a computing power of 700 TOPS, which can integrate up to 14 cameras and multiple high - speed data connections into a single "central brain". At the same time, it expanded its cooperation with Mahindra and launched the SmartCore Pro (triple - screen + 360 - degree view) for the upcoming XUV700.

To address the "legacy platform" obstacle, Visteon also launched a plug - in solution for the AI - ADAS Compute Module, powered by NVIDIA DRIVE AGX Orin, allowing automakers to add AI assistants or safety features without completely reconstructing the architecture; Deutsche Bank mentioned that this product has been installed in Geely's Zeekr models in China.

In addition, Visteon launched the "Entry Cockpit" for screens smaller than 7 inches, bringing mobile phone projection and digital navigation to two - wheeled vehicles and entry - level models. Deutsche Bank evaluated that its "surgical" vertical integration contributes to cost competitiveness and promotes further expansion among automakers with lower previous penetration rates (especially Asian OEMs).

In Deutsche Bank's view, the message conveyed by CES 2026 is straightforward: Autonomous driving and humanoid robots are moving from "can they be done" to "can they be scaled up and can the cost be reduced".

When Boston Dynamics pointed out that actuators account for about 60% of the cost and announced the pre - allocation of 2026 production, the industry has started to use manufacturing language for pricing; and Waymo's over 10 million paid rides and the rhythm of 450,000 rides per week are pushing Robotaxis from a concept to more tangible operational data. For investors, the next stage to track is not more dazzling demos but supply chain binding, production capacity ramp - up, and unit cost curves.

This article is from the WeChat official account "Hard AI", author: Long Yue, published by 36Kr with permission.