DEV Community

Shawn
Shawn

Posted on

FutureX · Physical AI Daily — Issue 64 (07/21)

Today's Highlights

· Blackstone invests about $676 million in South Korean actuator maker Futronic, betting on the robotics "components layer"

· China's Ministry of Industry and Information Technology: China has over 400 models of humanoid robots, about half the global total; quadrupeds account for nearly 70% of global sales

· Xiaomi releases Xiaomi-Robotics-1: a VLA model trained on 100,000 hours of real-robot data, pushing RoboCasa365 SOTA to 57.6%

· Yuequan Bionics (Chinese bionic robotics startup) raises a Pre-A+ round worth hundreds of millions of RMB; bionic dexterous hands reach 10,000-unit production scale with orders exceeding RMB 100 million

· UBTECH's humanoid robots to enter a South Korean auto-parts production line for a POC trial in August, marking embodied AI's overseas push shifting toward supply-chain integration

I. Research Progress

Xiaomi-Robotics-1: Scaling VLA with 100,000 Hours of Real-Robot Data · vla

Xiaomi has pushed VLA "data scaling" to a new magnitude — pretraining on 100,000 hours of real-robot manipulation trajectories (collected via UMI devices), paired with an automated annotation pipeline that labels each trajectory with natural-language descriptions of scene state transitions, providing fine-grained conditioning for action learning. The method splits into pretraining and post-training stages, showing a clear scaling curve: bigger models and more data yield better out-of-the-box real-robot performance, with the advantage transferring directly to post-training. It sets a new SOTA on RoboCasa365 with a 57.6% success rate (previous best: 46.6%), and averages 20.07 on RoboDojo (previous: 13.07); code and weights will be open-sourced.

Xiaomi Robotics Team · arXiv 2607.15330 source · Signal HF↑45

Harness VLA: Wrapping a Frozen VLA in a "Scheduling Shell" Instead of Retraining It · vla

A team led by Yu Chao at Tsinghua University proposes: rather than continuing to scale models and stack more robot trajectories, borrow from coding-agent design and give VLA its own harness. The frozen VLA is wrapped as a primitive handling only rich-contact actions like grasping and constrained placement, while contact-free movement, navigation, and release are handled by deterministic controllers; a planner then decides when to invoke the VLA, how to pre-position before invoking it, and how to retry after failures. With the same frozen VLA and unchanged parameters, simply changing how it is organized and invoked raised the success rate on the LIBERO-Pro generalization benchmark from 50.0% to 82.4% (compared with Nvidia's Cap-X at only 18.2% in the same period).

Yu Chao's team (Tsinghua University) · arXiv 2607.08448 source · Analysis: Embodied AI Insider source (WeChat, CN)

Orbis 2: Splitting Driving World Models into "Coarse Structure + Fine Frames" · world-model

Current world models mostly pursue frame-level fidelity at a single level of abstraction, lacking spatial and semantic reasoning. Orbis 2 splits future prediction into two levels: a high-level predictor forecasts coarse-grained scene structure over long horizons, and a low-level generator produces detailed frames based on that structure, balancing fidelity with reasoning. It also proposes a two-stage paradigm — "diffusion-forcing pretraining + teacher-forcing fine-tuning" — combining the representation quality of the former with the autoregressive stability of the latter, achieving SOTA across a suite of driving world-model benchmarks including long-horizon generation and counterfactual steering response.

Mittal et al. (University of Freiburg) · arXiv 2607.15898 source

EmArm: Giving Robotic Arms "Whole-Arm Human-Like Touch," Published in Nature Sensors · perception

By combining a rigid skeleton with flexible tactile skin, researchers built EmArm, a robotic arm with whole-arm human-like tactile perception that enables real-time physical adaptation — stably completing rich-contact manipulation while safely replanning trajectories in dynamic, unstructured environments. Compared with tactile sensing limited to the fingertips, it extends the "sensing zone" from the hand across the entire arm, pointing toward a contact-interaction approach closer to human touch.

Tang et al. · Nature Sensors source

Think at 5 Hz, Act at 20 Hz: Keeping LLM-Based Driving in Step with Control Frequency · autonomy

When large models are introduced into end-to-end driving, inference latency has always lagged behind the control frequency vehicles require, forcing existing approaches to invoke the model only every other cycle and reuse old commands in between — effectively ignoring the latest observation half the time. This paper proposes an asynchronous fast-slow architecture: a slow channel "thinks" at 5 Hz while a fast channel "acts" at 20 Hz, so every control step uses the latest frame. Open-loop waypoint error drops to nearly one-quarter of that from a backbone's built-in action head, with per-step model overhead of 32 ms independent of history length, runnable on a single consumer-grade GPU.

Li et al. · arXiv 2607.15621 source

Handroid: Combining "Dexterous Hand" and "Humanoid" into One Desktop Robot · manipulation

Dexterous hands and humanoids are typically developed separately — the former for object-scale rich-contact manipulation, the latter for locomotion and whole-body interaction. Handroid combines both into a single reconfigurable desktop-scale dual-morphology robot, uniformly supporting hand teleoperation, dexterous grasping, in-hand manipulation, and humanoid gait generation — providing a low-cost testbed for studying "hand × whole-body" coordination on one platform.

Li et al. · arXiv 2607.16187 source

Other papers today: AC-VLA uses compositional learning to strengthen out-of-distribution action execution, delivering roughly a 28% absolute improvement on compositional generalization tasks (arXiv 2607.15714 source); IMBench proposes an "intuitive manipulation" benchmark spanning perception—physical reasoning—action—iterative execution (arXiv 2607.15641 source); VTAP Gripper performs in-hand manipulation with a visuo-tactile active palm plus reconfigurable fingers (arXiv 2607.15448 source); MotionForesight adapts video models to predict future 3D scene flow from monocular video (arXiv 2607.16192 source); VTLoc uses 3D point clouds for tactile contact-point localization (arXiv 2607.16146 source); OptiTac uses fiber optics to mimic "artificial nerves" for bionic tactile skin (arXiv 2607.15746 source).

Open Source · Tools · Benchmarks

· Kuavo (X Humanoid): released what it calls the "world's first" large-scale cross-embodiment dataset and simulation engine, aimed at stable deployment of robots in real industrial scenarios source

· DexForce: open-sourced the Aligned DexWorld dataset, aligning dexterous manipulation data source

· Kunlun Wanwei (Chinese AI company) Matrix Game-3.5: open-sourced a game world model, focused on real-time responsiveness moving from "video generation" toward an "interactive world" source (WeChat, CN)

II. Funding & Deals

Futronic (South Korea) | Controlling-stake acquisition | About $676 million | Valuation of about KRW 1 trillion · hardware

Blackstone's private equity fund signed a definitive agreement to invest in Futronic, a South Korean maker of high-precision actuators and motion control systems, with founder Ko Jin-ho remaining as chairman and CEO — this is not a spin-off but a bet on expanded capacity. Founded in Busan in 1993, Futronic supplies actuators and motors for automotive and industrial robots, employs about 225 people, and is a company with three decades of experience in motion hardware. Rather than backing a front-facing humanoid brand, Blackstone is buying directly into the components layer that lets robots' "joints move and sustain load" — a so-called "sell the shovels" strategy: whichever full-robot maker wins, they will all need actuators.Source: Startup Fortune (citing Reuters/Bloomberg) source

Yuequan Bionics (Chinese bionic robotics startup) | Pre-A+ round | Hundreds of millions of RMB · hardware ⚠️ Order figures per company's own statement

Investors including Changfa, Huakong, Huaxia, Guoli Minsheng, and Xingke Venture Capital jointly invested, with existing shareholder Zhongguancun Qihang following on, marking the company's fourth funding round. Yuequan was founded by Ren Lüquan, an academician of the Chinese Academy of Sciences, and Ren Lei, a professor at the University of Manchester, pursuing a "bionic tension-compression structure" approach: self-developed magnetic-collector-drive artificial muscles integrate actuation, speed variation, and transmission into one unit, mimicking the human bone-muscle-tendon structure in an attempt to move beyond the conventional "motor + reducer + rigid joint" approach. Its flagship product, the Y-Hand M1 dexterous hand, has 38 degrees of freedom, full-hand tactile sensing, a grip strength of 28.7 kg, and single-finger repeat positioning accuracy of 0.04 mm. The company states its dexterous hands have reached 10,000-unit production scale with cumulative orders exceeding RMB 100 million.Source: 36Kr source (WeChat, CN) | Jingguan Ganzhi source (WeChat, CN)

Dobot (Chinese collaborative robotics maker) | Equity dispute alleged by co-founder ahead of ChiNext IPO hearing · industrial ⚠️ Dispute per the accusing co-founder's account

Collaborative robot maker Dobot faces an equity dispute alleged by a co-founder just ahead of its ChiNext IPO listing-committee hearing, with the company's flurry of recent funding rounds also drawing scrutiny. On the other hand, the company disclosed that first-half revenue and gross profit nearly doubled, with its collaborative-robot business becoming a "cash cow" underpinning its push into embodied AI commercialization.Source: Huxiu source

TÜV (Tuosida, Chinese industrial robot maker) files again for Hong Kong IPO, focused on integrated industrial robotics · industrial

Tuosida, whose core business is integrated industrial robots and self-developed R&D, has filed a second time for a Hong Kong listing, continuing the recent wave of Chinese industrial robotics companies rushing to list in Hong Kong during the current listing window.Source: 10jqka source

III. Commercialization & Deployment

UBTECH's Cruzr Y1 and Walker S2 to Enter a South Korean Auto-Parts Production Line · humanoid ⚠️ Still at POC validation stage

UBTECH signed a memorandum of cooperation with South Korea's Doksan JMR and AJIN Industrial to launch a production-line proof-of-concept (POC) for its Cruzr Y1 and Walker S2 humanoid robots at AJIN's newly opened manufacturing innovation factory in August 2026. Unlike its domestic-China deployments directly inside vehicle assembly plants for on-the-job training (Dongfeng Liuzhou Motor, Zeekr), this overseas move deliberately targets an auto-parts supplier, adopting a "UBTECH core technology + local Korean service" model to sidestep trade and compliance barriers. Whether the POC can convert into bulk orders remains to be seen, and local players like Rainbow Robotics are also accelerating their own efforts — this is one example of Chinese embodied AI's overseas expansion shifting from "exporting complete robots" to "localizing within the supply chain."Source: Netease Auto/Wangtongshe source

Jiushi Intelligence (Chinese autonomous-driving company): L4 Map-Free Autonomous Driving System Announces Mass Production · autonomy ⚠️ Per the company's own statement

At WAIC, Jiushi Intelligence stated its L4-level, HD-map-free autonomous driving system has entered mass production, emphasizing a low-cost, fast-deployment map-free approach.Source: Sina Finance source

YingShen's Embodied Robots Enter a Vietnamese Factory · industrial

YingShen's embodied robots have been deployed on a factory production line in Vietnam, marking another step in Chinese embodied AI's expansion into Southeast Asian manufacturing settings.Source: Vietnam Investment Review source

IV. Industry News

China's Ministry of Industry and Information Technology: China Has Over 400 Humanoid Robot Models, About Half the Global Total · humanoid

At a State Council Information Office press briefing on July 20, Wang Weiming, chief engineer of China's Ministry of Industry and Information Technology, shared first-half figures: China-developed quadruped robots account for nearly 70% of global sales share, and China has over 400 models of humanoid robots — more than half the global total. This follows Xi Jinping's appearance at WAIC setting the tone for "AI moving into the physical world" (previously reported), marking another official endorsement of the embodied AI industry's scale; WAIC also closed in Shanghai the same day. The same briefing also disclosed: China's new shipbuilding orders account for 81.2% of the global total, AI visual inspection has reached micron-level defect detection, and industrial large models have shortened preclinical R&D cycles for anti-tumor targeted drugs by 40%. The fact that over half of all model variants are Chinese-made indicates the "hardware supply" side is already highly saturated, with industry focus shifting from "whether it exists" to "who can deliver at scale and get robots actually working."Source: cnii.com.cn source (WeChat, CN) | Xinhua source

Unitree Releases UnifoLM-OminiA-0.3 Home-Care Embodied Model · embodied ⚠️ Per demo footage

Deployed on the G1, the model emphasizes a "single unified architecture," folding navigation, grasping, visual recognition, voice interaction, whole-home device coordination, and safety avoidance into one model — replacing the traditional layered architecture of separately calling perception/language/control APIs. In official demos, the G1 moves between living room, bedroom, and kitchen following a single spoken instruction, completing long-sequence tasks such as carrying a pillow, sorting dirty laundry, loading a dishwasher, and pulling a specific pill box out of a stack. The version number is deliberately marked 0.3, with technical details not yet disclosed; Unitree itself acknowledges the model is "somewhat slow to react, with occasionally clumsy grasping postures." This marks the fourth iteration in Unitree's UnifoLM series over the past year (world model → VLA → industrial → home care).Source: Deep Blue Embodied Intelligence source (WeChat, CN) | STCN source (WeChat, CN)

Fourier Intelligence (Chinese humanoid robotics company) Releases First Wheeled Dual-Arm Robot, GRW · industrial ⚠️ Per manufacturer specs

The brand's first wheeled-configuration robot targets flexible mobile assistance in factories — line-side delivery, tote handling, machine-side loading/unloading, palletizing and depalletizing, and other auxiliary stations that have long lacked automation. Specs: 29 degrees of freedom, dual-arm payload of 16 kg, peak torque of 500 N·m, repeat positioning accuracy of ±0.1 mm, a shoulder width of just 585mm to fit standard factory aisles, hot-swappable battery support for 24/7 continuous operation, and a no-code (natural language + drag-and-drop) tool chain that lowers the deployment threshold to accessory level. Fourier's reasoning: most factories have flat, hardened floors, so bipedal locomotion's terrain advantage goes unused while costing more in energy and stability — hence choosing a wheeled design to fill the automation gap at auxiliary stations first.Source: Robotics Grand Lecture Hall source

Hyundai Reportedly Moving to Further Tighten Control Over Boston Dynamics · humanoid

Following Hyundai's earlier $325 million acquisition of SoftBank's remaining 9.65% stake in Boston Dynamics (previously reported), South Korean media report that Hyundai Executive Chair Chung Euisun's personal stake in Boston Dynamics has risen to about 25%, as the group further consolidates control, paving the way for Boston Dynamics to be fully folded into Hyundai's robotics portfolio.Source: 매일경제 (Maeil Business Newspaper) source

Tesla to Report Q2 Earnings Wednesday, With Robotaxi and Optimus Progress in Focus · autonomy

The market is watching whether Tesla can demonstrate that its AI/Robotaxi/Optimus investments are "strengthening the company," with roughly 480,000 vehicles delivered in Q2. Separately, according to NHTSA data, Electrek reported that a Tesla remote operator was linked to a Robotaxi collision in Houston, adding further controversy over the safety of its remote-teleoperation setup.Source: Investor's Business Daily source | Electrek source

Hardware & Supply Chain

· Wolong Electric Drive × Shangwei New Materials: reached a partnership at WAIC combining motor drives with composite materials, entering upstream joints and structural components for humanoid robots source

· Solid-state batteries: industry discussion is increasingly framing their economics as a key variable for humanoid robots' battery life and safety source

Top comments (0)