Physical AI News (March 3, 2026 Issue)
Update Date: 2026/3/3
Executive Summary
As of March 2, 2026, a clear trend has emerged led by telecommunications carriers to utilize networks and GPU infrastructure as an "external brain" for robots. KDDI and AVITA are jointly developing a domestic humanoid, aiming for social implementation in customer service through on-premise Gemini integration. NTT DOCOMO and NTT have demonstrated low-latency remote GPU video analysis within safety requirements using IOWN APN and INCEdge. SoftBank has announced TelcoAICloud. Progress is also being made in Xiaomi's factory operations, Qualcomm and Siemens' autonomous factory, Toyota Tsusho's L2 public road demonstration, and AGIBOT's RaaS.
1️⃣ KDDI × AVITA: Strategic Partnership for Joint Development of Domestic Humanoid
Source: KDDI and AVITA Launch Strategic Business Partnership in the Physical AI Domain | KDDI News Room
On March 2, 2026, KDDI and AVITA launched a strategic business partnership in the physical AI domain. By combining AVITA's avatar production and remote control technology and domestic physical AI technology with KDDI's 5G and other communication infrastructure and GPU platforms, they aim to implement domestic humanoids in society that can handle tasks requiring non-verbal communication, such as customer service. A concept model assuming a customer service scene is scheduled to be exhibited at "MWC26 Barcelona," held in Spain from March 2 to 5, 2026. In addition, they are considering a configuration that links the GPU at the Osaka Sakai Data Center with Google's generative AI model "Gemini" on-premise, and plan to introduce it in commercial facilities and other locations on a trial basis from the autumn of 2026 onwards.
2️⃣ NTT DOCOMO / NTT: Successful Demonstration of 5G × Remote GPU Low-Latency AI Video Analytics
Source: DOCOMO and NTT Successfully Demonstrate Low-Latency AI Video Analytics | NTT Press Release
On March 2, 2026, NTT DOCOMO and NTT utilized the "IOWN All-Photonics Network (IOWN APN)" and "INC Edge" implemented on the 5G core network to demonstrate that the end-to-end latency of AI video analysis by remote GPUs falls within the threshold of ISO/TS 15066, which defines safety requirements for collaborative robots. This is a pioneering achievement that demonstrates the feasibility of remote AI inference support for autonomous robots in the 6G era through In-Network Computing (INC) technology, which integrates AI inference and communication control on the network side. They will also promote international standardization for device weight reduction, power saving, and decentralization in the future.
3️⃣ SoftBank "Telco AI Cloud" Concept: Communication Networks Become the External Nervous System for Robots
Source: SoftBank Corp. Announces Telco AI Cloud Vision | PR Newswire
At MWC Barcelona 2026, SoftBank announced its new vision, "Telco AI Cloud," to build next-generation social infrastructure for the AI era. It aims to realize decentralized AI infrastructure by integrating large-scale AI data centers (GPU cloud), AI-RAN-based MEC platforms, and its proprietary software stack, "Infrinia AI Cloud OS." Through a joint demonstration with Ericsson utilizing AI-RAN, they confirmed a low-latency, high-reliability network for Physical AI. They have also started collaboration with Mitsubishi Heavy Industries in the edge data center field and will promote the deployment of secure edge AI inference environments for industries such as manufacturing.
4️⃣ Xiaomi: Humanoid Achieves 90.2% Success Rate in EV Assembly Factory
Source: Xiaomi Humanoid Robot Hits 90.2% Success Rate in Auto Factory Pilot | ChinaEVHome
Xiaomi's humanoid robot achieved a 90.2% success rate in a self-tapping screw installation task at its own EV factory during three hours of continuous autonomous operation, while meeting the line's required cycle time of 76 seconds. By combining the VLA model "Xiaomi-Robotics-0" and the tactile feedback model "TacRefineNet," it overcame real-world noise such as magnetic interference and irregular component placement. This is a presentation of specific figures from an actual operating line, not a demo video.
5️⃣ Qualcomm × Siemens: Private 5G × Edge AI Autonomous Factory Demo at MWC
Source: Qualcomm Brings On-Premises Industrial AI to Siemens Factory Model at MWC | MarketScreener
On March 1, 2026, Qualcomm and Siemens jointly exhibited an autonomous factory model combining Private 5G and Edge AI (Cloud AI 100) at MWC Barcelona. AGVs and industrial robot arms operated in coordination, and AI agents on Siemens IPC performed real-time quality inspection and fault diagnosis. This is a showcase of next-generation manufacturing that demonstrated the smartization of manufacturing sites through closed-loop control.
6️⃣ Toyota Tsusho: Public Road Demonstration of Semi-Trailer Autonomous Truck (L2) Begins
Source: Public Road Demonstration for Practical Application of Semi-Trailer Autonomous Trucks Begins | PR TIMES (Toyota Tsusho)
A consortium including Toyota Tsusho and Robotruck began a public road driving demonstration with an L2 autonomous semi-trailer between Shizuoka and Aichi (approx. 90 km) on March 2. Safety verification, transport operation evaluation, and profitability verification are being conducted in parallel (demonstration period: 2/13–3/12). This is an important milestone for the social implementation of Physical AI to address the structural labor shortage in the logistics industry (the 2024 problem).
7️⃣ AGIBOT: Unveils Humanoid Lineup for Europe with Rental Support at MWC
Source: AGIBOT Unveils Full Humanoid Robot Lineup and Rental Service at MWC 2026 | Gizmochina
AGIBOT exhibited full-sized humanoids for reception and guidance, as well as quadruped robots, at MWC Barcelona, and officially launched its direct robot sales and rental (RaaS: Robot as a Service) business for the European market. Short-term rental services starting from €899 per day significantly lower the barriers to initial adoption. This approach is attracting attention as a way to accelerate the spread of robots through both hardware and business models.
Comprehensive Analysis
March 2, 2026, marks the shift of the main battlefield from "on-device AI" to "distributed AI." The demonstration that integrated control of remote GPU inference via INC met the latency requirements for collaborative robot safety is directly linked to lighter, more energy-efficient robots and reduced operating costs. On the other hand, KDDI's domestic humanoid and on-premise generative AI considerations reflect industrial needs that prioritize the sovereignty and security of on-site data. Furthermore, Xiaomi's actual line figures, Qualcomm and Siemens' closed-loop control, and AGIBOT's RaaS indicate that technical competition has expanded to "ease of adoption" and "business models."
Future Points of Interest
The focus of the KDDI x AVITA trial starting this autumn will be whether they can turn the naturalness of customer service, operational rates, and maintenance costs into KPIs while protecting personal information using on-premise Gemini.
NTT's INCEdge demonstration showed latency within safety requirements, but we should keep an eye on worst-case values during congestion, redundancy, power efficiency, and the battle for leadership in standardization.
SoftBank's Telco AI Cloud is a concept that bundles GPU cloud, AIRAN, and MEC; the developer-facing APIs and billing design will determine the speed of adoption.
Xiaomi's 90.2% success rate is strong, but generalization to product changes and component variations, continuous collection of tactile learning data, and safety certification during shutdowns will be the next hurdles.
The closed-loop autonomous factory by Qualcomm and Siemens presents a vision where edge AI handles everything from inspection to control, making the establishment of OT security and clear lines of responsibility essential.
The core of Toyota Tsusho's L2 semi-trailer public road demonstration is profitability, and the highlight will be how it is reflected in insurance premiums, regulatory operations, and implementation requirements for fleet management systems.
While AGIBOT's European RaaS lowers the barrier to entry, the design of accident compensation, maintenance structures, and software update SLAs will be the evaluation criteria that determine how 'easy to rent' it is.
