ACQUIRING TARGETS...
· DroidHunt news
China Mobile has fully open-sourced Open-RAIL, an asynchronous inference middleware that sits between robot "brains" — vision-language-action (VLA) and world-action models (WAM) — and robot bodies. Its core fix is for the mismatch between slow, irregular model inference and the fast, regular control loops robots need: a two-stage trajectory smoothing and fusing pipeline removes the resulting motion jitter while preserving the original policy's actions. Per the project site, Open-RAIL is already adapted to four-plus heterogeneous humanoids (Unitree G1, an AgiBot platform, China Mobile's own Lingxi, and a Zhejiang humanoid) and supports 20-plus mainstream models (GR00T, TAO, RDT, DreamZero, PI, ACT). Integrating a new model takes 50–100 lines of code via a standardized interface; a server-client architecture lets inference run on-body, at the edge, or in the cloud with no upper-logic changes. It also records real-robot execution data at runtime and supports human-in-the-loop teleop intervention, feeding closed-loop reinforcement learning. China Mobile describes it as the industry's first general engineering base connecting VLA/WAM models to real robots.
Why it matters: Every humanoid team today rebuilds the same deployment plumbing between model and motor. A neutral, open-source execution layer — from a carrier with its own humanoid fleet ambitions — could commoditize that plumbing the way ROS did for robot software generally, and it gives China's ecosystem a shared rail at the exact moment VLA models are racing to real hardware.
Sources: forbes.com cmcc-tao.github.io en.tmtpost.com
Mech-Mind Robotics, the industrial 3D-vision and robot-guidance supplier, released its first interim report since listing. H1 2026 revenue was ¥237M, +54.7% YoY; new orders ¥335M, +75.3%; overseas revenue ~¥100M, +68.3%. Gross profit rose 63.6% to ¥154M, with gross margin up from 61.4% to 65.0%. Losses continue: operating loss widened to ¥77.4M as R&D spend jumped 72.9% to ¥80.1M (33.8% of revenue), though adjusted loss narrowed ~5.9% to ¥52.5M. Repeat-purchase rate was 93.9%, with 100+ Fortune Global 500 customers across ~50 countries; sales-expense ratio fell from 102.7% in 2023 to 38.5%. The report lands weeks after founder Shao Tianlan publicly questioned the revenue authenticity of some embodied-AI peers — naming Galaxy General in Sept 9–10 posts — and Mech-Mind's own books show the embodied-AI story is still ahead of its revenue: 93% of 2025 revenue (¥389M) came from its mature robot-guidance business, while the newer Mech-GPT multimodal model and Mech-Hand dexterous hand contribute little and are slated for scale launch only at end-2026, with meaningful commercial revenue expected from H2 2027.
Why it matters: The man who called out "fake revenue" in embodied AI now has his own numbers public — solid industrial-vision fundamentals, but zero proven embodied-AI product revenue yet. It's also a read-through for the 40+ embodied firms in the IPO queue: public markets will see exactly how much of "embodied intelligence" is booked sales versus narrative.
Sources: cls.cn
The IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) opens Sunday, Sept 27 at Pittsburgh's David L. Lawrence Convention Center and runs through Oct 1. Pre-announced demos relevant to the humanoid beat: Viam will debut BoxBot, a research arm that opens taped cardboard boxes — image-based visual servoing cuts the 1–2 mm tape seam, while a fine-tuned open-source SmolVLA policy trained on just 125 teleoperated demonstrations collected in about five hours handles contact-rich flap opening, with inference running locally on a Jetson Orin Nano; the implementation is public. Xsens will preview an upcoming integration with NVIDIA Isaac Teleop for full-body humanoid teleoperation (booth #721, with Robo Inc.) — an early look, not a shipping product. ROBOTIS will run live demos of AI Sapiens, its open-source humanoid platform (23 Dynamixel actuators). The "Athletic Humanoid Robotics" workshop features demos by Unitree, Phybot and Booster, with speakers including Tsinghua's Mingguo Zhao, CMU's Guanya Shi and Sony AI's Peter Stone. PAL Robotics brings its new TIAGo Pro Arm; Flexiv demos its force-controlled Enlight and a MICO robot installing GPUs into server racks.
Why it matters: IROS is where lab work meets deployment claims. The through-line this year is data efficiency and teleoperation infrastructure — BoxBot's 125-demo training run and the Xsens/Isaac Teleop pipeline both attack the demonstration-data bottleneck that every humanoid foundation-model team is spending heavily to solve.
Sources: prnewswire.com instagram.com iros-2026-athletic-humanoid.github.io
Tags: humanoid iros-2026 teleoperation vla
Tokyo startup O-ID has raised $1.2 million to develop modular industrial humanoids with swappable joints, limbs and compute units, designed so on-site repairs take minutes instead of days. Its first target markets are Japanese automotive manufacturing and logistics. The modularity thesis attacks a real cost center: humanoid downtime — if a failed actuator or joint can be swapped line-side in minutes, fleet operators avoid shipping units back for service, a meaningful factor in total cost of ownership for early industrial deployments. Caveats: a $1.2M round is seed-pocket-money in humanoid terms, and O-ID has shown no public hardware yet — treat this as an early-stage bet on a design doctrine rather than a product launch.
Why it matters: Modular, field-repairable hardware is a different design philosophy from the sealed, single-vendor-service approach of most current humanoids, and directly targets the downtime economics that will make or break early fleet deployments.
Sources: humanoidsdaily.com humanoid.guide
← Newer: Sep 28, 2026 Older: Sep 25, 2026 →
Home Every digest Robot catalog Drones on DronePriceHunt Privacy & affiliate disclosure RSS