Physical AI News (July 2, 2026)

Update Date: 2026/7/2

Executive Summary
On July 1, 2026, the shift of Physical AI from the research phase to commercial implementation, mass production, and operational infrastructure became clear. UBTECH announced the U1 humanoid for home and commercial use, and Apptronik built a real-world data collection loop for Apollo 2 at Robot Park. AGIBOT unveiled the A3 in the UK and began RaaS deployment. In addition, the implementation ecosystem is expanding across hardware, AI, data, and operations layers, including Tenstorrent's RISC-V CPU IP, tactile datasets, 3D trajectory guidance, inference latency reduction, the domestic Physical AI Exhibition, NRI's Physical AX, and SICK's field sensors.

*The created article content was input into Gamma to automatically generate slides. Please take a look here if you find slides easier to read.

1️⃣ UBTECH 'UWORLD U1': Announced mass-production design humanoid for home and commercial use

📎 Source: KYODO NEWS PRWIRE / PR Newswire / Moomoo
UBTECH held the '2026 Global Launch Event' in Shenzhen on June 30, 2026, and officially announced the 'U1' series, a full-size super-bionic humanoid under the consumer brand UWORLD [page:1]. The lineup consists of three models: U1 Lite (semi-torso), U1 Pro (full-body high-performance), and U1 Ultra (full-body high-dynamic), with prices starting from 119,800 RMB [page:1]. The cumulative number of orders at the time of the announcement was 'over 10,000 units,' and the press release posted on Kyodo News PR Wire stated that it exceeded 13,361 units [page:1]. The U1 series features 88 degrees of freedom and a proprietary dual-pivot biomimetic cervical spine. It utilizes an LLM specialized in emotion recognition to recognize over 20 granular emotional states with over 90% accuracy, and features a cognitive architecture that combines 500ms intuitive responses with deep reasoning from large-scale models [page:1]. It also promotes a three-layer privacy design consisting of local-first processing, minimization of cloud dependency, and user-controlled hardware protection, targeting both consumer and commercial applications such as daily companionship and emotional support at home, elderly care, reception and customer service, psychological support, tourism and exhibitions, research and education, and premium household services [page:1].

2️⃣ Apptronik 'Robot Park': Real-world data loop for Apollo 2 and Gemini Robotics

📎 Source: Apptronik / Reuters
On June 30, 2026, Apptronik opened 'Robot Park,' a humanoid training facility expanded to approximately 90,000 square feet in Austin, Texas, and announced its latest humanoid platform, 'Apollo 2.' Both bipedal and wheeled versions of Apollo 2 perform tasks in logistics, manufacturing, and retail by combining remote operation and autonomous movement, continuously collecting large-scale real-world data. This data is used for training Gemini Robotics and developing the next-generation commercial machine, Apollo 3, in research collaboration with Google DeepMind.

3️⃣ Tenstorrent 'TT-Ascalon S': RISC-V CPU IP for Agent AI and expansion into Japan

📎 Source: Tenstorrent / PC Watch
On June 30, 2026, at the 'TT-Deploy JP' event held in Tokyo, Tenstorrent announced 'TT-Ascalon S,' a high-compute-density RISC-V CPU IP for agent-based AI. Based on the TT-Ascalon X, it is designed to reduce the footprint to approximately 50% while aiming for about 140% performance per square millimeter, optimized for orchestration, I/O, and low-latency processing handled by the CPU. At the same time, the company introduced a domestic sovereign AI inference infrastructure using Tenstorrent Galaxy and a PoC by autonomous driving startup Turing using Blackhole in vehicles, emphasizing deployment for edge inference applications such as agent-based AI, robotics, and autonomous driving.

4️⃣ AGIBOT A3: European debut at UK APC2026, RaaS deployment begins

📎 Source: AGIBOT
AGIBOT held the UK AGIBOT Partner Conference 2026 in London on June 30, 2026, and unveiled the new generation humanoid 'AGIBOT A3' in Europe for the first time. The A3 features a lightweight body of 173cm height and 55kg weight using a reinforced structure of magnesium-titanium alloy and TPU, up to 10 hours of continuous operation, hot-swappable battery replacement in about 10 seconds, and centimeter-level positioning and multi-robot coordination via UWB. The company announced an RaaS model for the UK market, offering humanoids from £1,999 per day and quadruped robots from £899, indicating a policy to accelerate commercial deployment in Europe.

5️⃣ RoboTacDex / RCT: Robot learning datasets including tactile sensing released one after another

📎 Source: arXiv:2606.31836 / arXiv:2606.31694
RoboTacDex, submitted on June 30, 2026, is a humanoid manipulation dataset using the Unitree G1, containing 6,000 trajectories, 19 tasks, 23 skills, and interactions with 22 types of objects, integrally recording multi-view RGB, depth, tactile, and detailed semantic annotations. Meanwhile, RCT is a multimodal dataset of tactile, visual, and language data collected with three DIGIT sensors for 29,279 tactile frames across 122 types of industrial reference materials, focusing on material property recognition and generalization. Both indicate a trend where the center of gravity of robot learning is expanding from image and language-centric to rich physical interactions including contact, force sensing, and material generalization.

6️⃣ 3D HAMSTER / ELASTIC: 3D trajectory guidance and inference-time computational optimization

📎 Source: arXiv:2606.31329 / arXiv:2606.31132
3D HAMSTER proposed a framework where the VLM side directly generates geometrically consistent 3D trajectories in meters rather than 2D paths, as a bridge between high-level planning and low-level control in hierarchical VLAs. By introducing a dedicated depth encoder and dense depth reconstruction objectives, and integrating predicted 3D waypoint sequences into a point-cloud-based low-level policy, it consistently outperforms existing VLMs and 2D guidance baselines under appearance changes and unknown language, spatial, and visual conditions. ELASTIC is a framework that elastically adjusts the inference-time computational cost of generative control policies per state, and it is reported to have reduced end-to-end latency by approximately 34% while maintaining almost the same success rate in the real-world π0.5 VLA.

7️⃣ Physical AI Exhibition Tokyo / NRI Physical AX: Domestic implementation ecosystem expands

📎 Source: Physical AI Exhibition / NRI
From July 1 to 3, 2026, the '1st Physical AI Exhibition Tokyo' opened at Tokyo Big Sight, covering a wide range of Physical AI implementation themes for the manufacturing industry, such as AI robotics, operational management, safety and communication infrastructure, control AI, sensing, and simulation infrastructure. On the same day, Nomura Research Institute (NRI) announced the promotion of 'Physical AX' through its participation as a regular member of the AI Robot Association (AIRoA), indicating a policy to advance the digitization of tacit knowledge dormant in the field, the implementation of an AX platform that consistently integrates AI and physical devices, and the construction of a cross-industry DX and robot data ecosystem.

8️⃣ SICK 'LUTS/LUTX': Enhancing field perception with fluorescent marker detection sensors

📎 Source: AutomationNews / SICK Product Page
SICK has released the "LUTS/LUTX" luminescence sensors, which reliably detect visible and invisible luminescent markings. Equipped with UV LED (375nm) or blue light LED (470nm) emission variations, they can stably detect fluorescent markers on dirty, glossy, transparent, and highly reflective materials. Available in two sizes—the compact LUTS for tight installation environments and the larger LUTX for flexible installation—they feature an integrated display showing luminescence intensity values from 0 to 999, a switching frequency of up to 16kHz, scalable digital output with a factor of 0.5 to 4, and support for diagnostics and rapid format changes via IO-Link. They contribute to improving recognition stability for robots and inspection equipment in industries such as packaging, pharmaceuticals, and wood processing.

Comprehensive Analysis

The key trend observed from the topics on July 1, 2026, is that the competitive axis for humanoids and robot AI has shifted from "moving demos" to "mass production design, field data, low-latency inference, and deployment models." UBTECH and AGIBOT have prioritized deployment conditions such as price, operating time, and RaaS, while Apptronik has established facilities to continuously collect real-world data. In research, advancements in tactile sensing, material recognition, 3D trajectory planning, and computational load control are laying the foundation for AI that is less prone to failure in the physical world. Domestically, through exhibitions and NRI's participation, the trend of digitizing tacit knowledge from manufacturing sites and connecting it to robot operations is strengthening.

Future Points of Interest

• For humanoids, evaluation metrics will shift from just reservation numbers and performance announcements to actual operating rates, maintenance costs, failure rates, and field retention rates.

• Real-world data collection facilities like Apptronik's Robot Park will become critical infrastructure that creates performance gaps between VLA and Gemini Robotics-based models.

• The expansion of tactile datasets will mark a turning point from robots that "see and judge" to those that understand materials and contact states to perform operations.

• Tenstorrent's RISC-V deployment is attracting attention as a domestic option for AI infrastructure in edge inference, autonomous driving, and robotics that reduces reliance on the cloud.

• With the spread of RaaS models, companies are likely to shift to an adoption style where they rent robots by the task for verification rather than purchasing them.

• In the domestic Physical AX initiative, the key will be how to standardize the tacit knowledge of manufacturing sites and reuse it as robot learning data.