Yu Hao proposes that evaluating on-device large-model chips should shift from TOPS to comprehensive metrics such as effective memory bandwidth, energy efficiency, and tokens/s. A 3D-DRAM stacked cubic systolic TPU architecture measured about 80 tokens/s for INT4-quantized Decode of a 3B model at an average power consumption of 5W, with token energy efficiency of 15 token/W and bandwidth utilization of about 80%. The related 3D-TPU solution is at the sample stage, and Maitexin's AI agent MetaClaw has been adapted to Qwen, MiniCPM, and Gemma and is undergoing pilot applications.
The article reviews the rise of the four major industrial robot families from the 1970s to the 2000s: KUKA launched the six-axis, all-electric KR FAMULUS in 1973; ASEA launched the IRB 6, combining all-electric drive with microprocessor control, in 1974; FANUC started out in NC and servos and mass-produced the ROBOT MODEL 1 in 1977; and Yaskawa launched the MOTOMAN-L10 in 1977.
A Forbes article explores how Hexagon Robotics is moving its humanoid robot AEON from single-unit tasks toward industrial deployment at a scale of 1,000 units. AEON uses a wheel-legged design that combines efficient mobility with obstacle-crossing capability, and is positioned as a multi-purpose industrial platform supporting manipulation, inspection, and real-world data collection. Hexagon Robotics uses a sim-to-real-to-sim closed loop to collect real-world data and feed it back into training, and cites BMW Group and Schaeffler as examples of the path from pilots to repeatable performance and industrial-scale deployment.
Drawing on experience with first-person robot data processing, the author points out that storing only RGB frames makes observations hard to localize in 3D space; camera intrinsics, distortion parameters, and stereo extrinsics, combined with matching algorithms, can recover depth in real units downstream. In its egocentric-view collection setup, the team synchronously records calibrated stereo video, camera poses, body skeleton, and hand coordinates, and combines this with motion tracking to support downstream tasks such as reconstruction, multi-sensor alignment, and cross-embodiment transfer. The author stresses that sensors and dataset design should be planned together, since calibration and synchronization errors can make the lowest-cost datasets hard to use.
The article analyzes how Wang Xingxing, at the China International Industry Fair, shortened his most optimistic estimate for the tipping point of embodied intelligence from two to three years to one to two years, and admitted he had not believed in AI enough. Reports from Phoenix News Technology and others say Liang Wenfeng has been meeting intensively with embodied intelligence entrepreneurs; DeepSeek has already made a strategic investment in Unitree and plans to lead a funding round for Winverse AI, raising tens of millions of dollars at a post-money valuation of about 400 million dollars.
OFweek published an analysis of which companies in the embodied intelligence industry might possess Apple-style product definition and integrated hardware-software capabilities, selecting Sharpa, Robot Era, Astribot, and SWancorp as samples representing four directions.
The author summarizes three investigations published on the same day by Caijing, Lanjinger, and Forbes, noting that the embodied AI industry has engaged in practices such as pass-through order transactions and inflated factory deployment narratives.
At the 26th China International Industry Fair, Unitree founder Wang Xingxing said his biggest mistake over the past decade was not believing in AI enough. He said the release of GPT-6 greatly improved embodied intelligence practitioners' understanding, and that Unitree tested it on various robots with good results. He believes a breakthrough could come within one to two years if things move quickly, and that GPT-7 next year could directly solve the embodied intelligence problem.
A highlight reel only proves a robot can perform a move once for the camera.
Deployment proves it can work safely on a real floor, shift after shift without interruption.
Only one of the two can support a scaled business.
Agility Chief Business Officer Daniel Diez discusses what truly distinguishes this market: http://www.agilityrobotics.com/content/scaling-a-new-market-category
#Agility #MadeForWork #Digit5
The higher the level of warehouse automation, the greater the impact of a single equipment failure, making uptime a system-level issue. The article notes that even highly automated facilities still rely on mixed fleets of traditional equipment such as forklifts and pallet trucks, and that automation has not reduced the importance of equipment maintenance. Preventive maintenance using telematics data (such as operating hours, fault codes, and utilization) can shift maintenance from reactive response to proactive management. Black Equipment is cited as an example of such an uptime strategy.
Productivity gains in warehouses and distribution centers are often limited by the "reality gap" between what digital systems perceive and what is actually happening on the warehouse floor, rather than by a lack of technology itself. The article argues that the focus should shift from automation to orchestration and real-time execution: continuously sensing operational status, detecting anomalies immediately, and dynamically allocating work and resources. Organizations that successfully close the gap can achieve greater resilience and more balanced load utilization, with competitive advantage shifting from assets to agility.
Productivity bottlenecks in warehousing and distribution often stem not from a lack of technology but from the 'reality gap' between what digital systems perceive and what is actually happening on the warehouse floor—information arrives too late, is trapped in silos, and lacks context. The article argues that while multiple automation subsystems may each work well, they lack coordination; the focus should shift from automation to orchestration, using real-time execution to continuously sense operating conditions, detect anomalies instantly, and dynamically allocate work and resources. Closing this gap can yield greater resilience to disruption, better resource utilization, and faster decision-making, with competitive advantage shifting from assets to agility.
A robotics developer asks OEMs and machine builders: when facing multiple product specifications, should they keep mechanical structures such as guide rails and feeders to ensure determinism, or simplify the mechanics and hand the variability to robot-plus-vision software?
The author analyzes the real-world distance to robotic elder care, noting that caregiving involves three types of labor—physical, emotional, and cognitive. Robots currently excel at information-based tasks such as monitoring, reminders, and conversation, while the most anticipated physical labor remains the biggest bottleneck due to a lack of real interaction data. Drawing on 20 years of experience with Japanese care robots, the article argues that small, specialized vertical products are more likely to survive, while general-purpose care robots mostly remain prototypes. It derives an implementation sequence: institutions first, then home; commercial services first, then household consumer use.
Eliana Otrera, a teacher in Argentina's Santiago del Estero province, brings programming and robotics education to rural schools through Raspberry Jams. On March 6, 2020, she organized the province's first community-oriented Raspberry Pi workshop in Forres, and has continued holding them at schools, cultural centers, and institutes.
Wai-keung Fung of Cardiff Metropolitan University writes that at the World Humanoid Robot Games 2026, Tiangong Ultra completed the 100 meters in 8.64 seconds, faster than Usain Bolt's 9.58-second world record, but soft manipulation tasks such as folding clothes and doing laundry are a better test of robot intelligence. The article notes that fabric shapes vary widely and skills are hard to transfer directly between robots with different configurations and sensors, and that Open X-Embodiment is exploring this problem by gathering over a million real trajectories from 22 kinds of robots.
Vishay Precision Group innovation director Ron Zukerman argues in an article that robot design should work backward from economically valuable tasks to hardware, perception, and form, rather than building a platform first and then looking for tasks.
This is a community discussion post asking why even the most advanced humanoid robots, which are still controlled by large-scale digital neural networks, make mistakes such as dropping cups or crushing eggs due to latency, and questioning whether digital computing power has fundamental limitations for dexterous hands. The poster points out that biological reflex circuits are analog and asynchronous, and asks how many companies are developing mixed-signal or neuromorphic hardware for robotic hands rather than simply scaling up software.
TDK Ventures founder Nicolas Sauvage argues in an article that the key to robotic dexterity is not completing a single action, but the reliability of an entire workflow. He gives the example that when the single-step success rate is 99%, the probability of completing all 100 actions is only 36.6%, and only when it rises to 99.99% does it reach 99.0%.
Robotics & Automation News interviews Steve Ricketts, senior vice president of business development at Misumi Americas, about the manufacturing challenges of taking humanoid robots from prototype to mass production. He sees actuators as the biggest bottleneck, with limited capacity for high-precision gear transmissions, rare-earth magnets, and five-axis machining. Cost reduction will mainly come from design simplification, component standardization, and a shift from machining to castings and injection molding. He expects common parts such as motors and bearings to become standardized, while software, hand design, and system architecture remain proprietary.
In episode 265 of The Robot Report Podcast, Jeanine Sinanan-Singh, director of generative AI research at Appen, explains why impressive robot demonstrations do not necessarily prove that their behaviors can generalize, and emphasizes that human-in-the-loop training of robots remains important.
Boston Dynamics introduced its redesigned four-finger dexterous hand for the humanoid robot Atlas, featuring 13 degrees of freedom, direct drive, the ability to handle objects over 100 pounds, and a design intended for mass production. Director Alberto Rodriguez said the hand drops the pinky finger, uses 13 fully enclosed identical actuators with no fragile cables spanning joints, and is designed for high-fidelity simulation and sim-to-real reinforcement learning; the team has tested industrial tasks such as power tool use, cable handling, and screw finding, but has not yet deployed it across the entire robot fleet.
In this episode of the podcast, Claire talks with Kushant Patel of Brookhaven National Laboratory about robots that work in scientific facilities such as particle accelerators. Patel develops software that helps robots and instruments perceive, map, and move through the physical world. He previously taught quadruped robots to map and inspect nuclear materials at UC Berkeley, and now develops control systems for advanced synchrotron light beamlines at Brookhaven.
Xinhua News Agency published a hot-topic briefing titled "What is the biggest difference between humanoid and traditional robots?", which addresses hot topics around humanoid robots by drawing on the views of Wei Hongxing, founder and chairman of AUBO (Beijing) Intelligent Technology Co., Ltd. The original article does not provide specific technical indicators or product parameters; it only republishes the core content of the report so readers can follow the pulse of the humanoid robot industry's development.
No training needed, no sim-to-real gap, just a prompt and a vision.
Sentdex points out that every task in robotics today is a complex combination of locomotion, manipulation, vision, planning, and more, and fails once it deviates from training data or simulation characteristics. As a result, many demos are conducted only under perfect lighting and carefully placed objects, and never land under unknown real-world conditions—every deployment is a unique edge case.
In the example, a general-purpose multimodal large language model (DSV4.1F) completed a robot task it was never specifically trained for, using an unseen camera, unlearned servo motors/motors, an environment it had never handled, and a morphology it had never been tested on. The video is at 10x speed and is currently slow, but this is already the slowest state at this level of intelligence.
Sentdex says that the more he uses general-purpose multimodal large models, the more he realizes that specialized methods such as ACT and VLA may be mere supporting acts and distractions, and will eventually be swallowed by general models that do robotics incidentally. He does not think it will always be large language models, but the fact that large models can complete the most challenging robot tasks shows that...
—Sourccey functionality is continuously improving.
Most industries treat safety as a cost to clear before shipping.
We see it as the key to opening the next market—each layer larger than the last.
Why safety is strategy: https://www.agilityrobotics.com/content/built-for-the-real-world
Thomas Brown of manufacturing services provider Jabil spoke with The Robot Report about the cost structure and manufacturability required for humanoid robots to reach the tens-of-thousands scale. He said edge compute and price remain pain points, that consumer-grade robots may arrive in about five years, and that humanoids will not replace traditional high-speed automation; safety certification and failure-mode research still require substantial work.
AI has already learned to generate answers in the digital world. The next step is taking action in the physical world. At #APRCE2026, Zhengxing Innovation founder and CEO Song Yao shared how AI innovation is reshaping 7×24 #retail and bringing convenience to the world through #physicalAI.
Nvidia is expanding its Halos operating system from autonomous driving to robotics, allowing developers to customize safety features without disrupting the underlying architecture. The system continuously monitors hardware and software libraries; Holoscan Sensor Bridge can be embedded in microcontrollers or FPGAs to identify corrupted sensor data. It also includes virtual environment simulation and an inspection lab program, providing safety feedback for robot development.
Starting from the caregiving experience of a robotics company founder, the article breaks elder care down into three tasks of differing difficulty: physical labor, emotional labor, and cognitive labor. It points out that what robots currently do well—monitoring, reminders, and conversational companionship—belongs to the information world, whereas the physical labor people most hope to have replaced requires large amounts of real-world data on human bodies, force, and contact, which is the industry's biggest bottleneck.
Leiphone's Zuo Lin You Li channel interviewed Benewake founder Li Yuan, reviewing his transition from optoelectronics scholar to entrepreneur. In the first half of 2025, Benewake delivered 250,000 LiDAR units for robots, with robotics business revenue accounting for more than half of the company's revenue; this revenue grew 100% in 2025 compared with 2024.
Agentic Robotics — using AI coding agents to design and fine-tune robot control systems with harness frameworks like ENPIRE, ASPIRE, or GaP — is very exciting.
If you know a small or medium-sized manufacturer applying Agentic Robotics in actual production to improve real manufacturing, please get in touch with me.
Last month, the author took part in the UK RAS STEPS MicroMaze Robot Hackathon (August 25-27) at the Institute for Safe Autonomy at the University of York. Together with teammates, the author used a laser-cut chassis, sensors, motors, a Raspberry Pi Pico W, and MicroPython to build a small robot capable of autonomously navigating a maze within three days. Beyond the technical work, the author made new friends and thanked the organizers for creating opportunities for research technical staff.
Boston Dynamics announced the appointment of Rohit Prasad as its new CEO, effective October 7, 2026. Previously, Robert Playter stepped down at the end of February 2026, and Amanda McMaster served as interim CEO. Prasad spent 12 years at Amazon, where he was head of Alexa and Artificial General Intelligence and a senior vice president, and led development of the Amazon Nova family of foundation models; the company will draw on his AI background to accelerate the commercialization of physical AI.
Network-level security and robot-hardware-level safety are two separate layers — and on Digit they stay that way.
Our CTO Pras Velagapudi explains why this is one of the most important design decisions behind deploying humanoid robots 👇
Watch the full breakdown: https://www.youtube.com/watch?v=WpGs7DKe8QY
#Agility #DataSecurity #Digit #BuiltForWork
Professor Enrica Tricomi of TUM MIRMI has established a new research group focused on neurorobotics and intelligent assistance, studying how wearable robots can understand and adapt to their users. She focuses on hybrid systems combining robotics with neuromodulation, acting on nerves and muscles through mechanical assistance and electrical or sensory stimulation to provide immediate movement assistance and help people with movement disorders (such as multiple sclerosis and Parkinson's disease) recover motor function over the long term. The lab is recruiting researchers with backgrounds in robotics, mechanical, electrical, and biomedical engineering, as well as neuroscience and rehabilitation.
Agility Chief Business Officer Daniel Diez discusses why real deployment—not demo videos—is the key to scaling this market: https://www.agilityrobotics.com/content/scaling-a-new-market-category
#Agility #MadeForWork #Digit5