Today's Highlights
· Pony.ai's Shenzhen robotaxis turn a monthly per-vehicle profit, averaging 338 RMB in daily net income
· Uber commits $1.25 billion, plans to deploy 50,000 Rivian robotaxis starting 2028
· China's industrial robot output hit 538,000 units in H1, up 28.0% year-on-year
· San Francisco humanoid home-cleaning service launches at $30/hour — actually human-operated via teleoperation control room
· Tsinghua's 1.8-gram microrobot can switch between quadruped and humanoid forms
I. Research Progress
A 1.8-gram robot that reshapes itself between quadruped and humanoid form · locomotion
At the centimeter scale, roboticists have long had to choose between reconfigurable form and fast dynamic response — machines that can transform tend to move slowly, while fast ones have fixed shapes. Tsinghua University's Zhang Yihui group connected a reconfigurable "skeleton" module with a fast-response, variable-stiffness "muscle" module in series, decoupling shape and dynamics for the first time: as the shape-memory polymer softens under heat, the muscle module's resonant frequency drops from 64 Hz to 14 Hz, and its bending angle expands from about 2.5° to 75°. Two prototypes resulted: a 1.8 g robot that switches between quadruped and humanoid configurations — running at 3 body-lengths per second and carrying loads over 3 times its own weight in quadruped mode, then using its front limbs for grasping and throwing once switched to humanoid form; and a 6.4 g wireless robot modeled on the sugar glider, which uses a 1.04 g boost-converter circuit to step a 3.7 V battery up to roughly 9.5 kV, freeing it from external power and letting it crawl on land, glide across water, and glide through the air.
Xu Shiwei et al. (Tsinghua University) · Science Advances · via DeepTech深科技 source (WeChat, CN)
ergoCub: designing the robot body around human lumbar load, then adding control · locomotion
The usual humanoid robot approach is to fix dimensions, weight, and linkages first, then tune the controller; this Nature Machine Intelligence paper reverses that order, coupling a human multi-body dynamics model with the robot's own dynamics into a single model, and using a dual objective — human lower-back torque plus robot center-of-mass height — to work backward to linkage dimensions. Iterating from an iCub prototype, the team arrived at the 150 cm, 56.7 kg ergoCub platform, then fixed the hardware parameters to optimize a three-layer, center-of-mass-MPC-based controller. In collaborative-carrying experiments, peak human L5-S1 lumbar torque dropped significantly, and the robot could track a human hand moving through a 0.8–1.2 m range with centimeter-level tracking error; carrying a 6 kg load, it could still adjust its footing to maintain balance under 60–100 N lateral pushes. The authors have open-sourced both the modeling and control framework.
Istituto Italiano di Tecnologia (IIT) · Nature Machine Intelligence · DOI 10.1038/s42256-026-01272-2 · via 人工智能与机器人进展 source (WeChat, CN)
MaskVLA: deliberately blocking the main camera during training makes VLA more robust · vla
VLA models fine-tuned on small datasets tend to "memorize the camera viewpoint" rather than understand the task — the model ends up memorizing local textures and spatial positions from the training demonstrations, then breaks down the moment an occlusion appears or an object moves. This work's approach is minimal: during training, the main-view image is randomly masked with a certain probability, forcing the model to rely on close-range information from the left and right wrist cameras, while full multi-camera input is still fed in at inference time. On RoboTwin 2.0, average success rate improved by 23.2% and 16.8% over π0 and OpenVLA-OFT respectively, reaching 75.0% on Hard mode; on a real ALOHA dual-arm setup, it likewise led on tasks like feeding carrots and stacking bowls. The method is plug-and-play and doesn't alter the base model's inference structure.
Wang Ge et al. (The Chinese University of Hong Kong, Shenzhen) · Accepted at IROS 2026 · via 港中大深圳理工学院理学硕士项目 source (WeChat, CN)
Open Source · Tools · Benchmarks
· Project Eden (VAST AI Research): a persistent world-model research preview from the team behind Tripo, splitting "world state" from "rendering" into a three-layer architecture — once a fire is put out it stays out, and a wall stays put even when the camera turns away; it supports multiple people entering the same world from different viewpoints, directly targeting the longstanding problem of frame-by-frame video generation losing context source
II. Funding & Deals
Uber × Rivian | Robotaxi Supply and Investment Agreement | $1.25 Billion · autonomy
Uber will invest $1.25 billion in Rivian by 2031 to support the deployment of up to 50,000 fully autonomous Rivian R2 robotaxis, operating exclusively on the Uber platform. The rollout begins with 10,000 vehicles in San Francisco and Miami in 2028, expanding by 2031 to roughly 25 cities across the US, Canada, and Europe. The R2s built for Uber will carry more sensors and dedicated software than the consumer version, plus redundant steering, braking, and upgraded thermal management, aimed at near-all-weather operation. Rivian's own roadmap calls for point-to-point driver assistance by late 2026, highway L3 by 2027, and L4 service starting in 2028 — meaning the fleet-size figures all sit downstream of capabilities that still need to be delivered.Source: 汽车之家 source
NXP × Ambarella | Acquisition Talks | Target Valued at ~$3.3 Billion · hardware ⚠️ Media report, unconfirmed
According to the Financial Times, NXP Semiconductors is in talks to acquire Ambarella, maker of edge AI vision chips for autonomous driving and security cameras, whose CV3 domain-controller chip can connect to up to 12 cameras. On the news, Ambarella's stock surged nearly 20% at one point while NXP fell more than 3%. One caveat: the $3.3 billion figure is Ambarella's market valuation, not a disclosed deal price — talks are ongoing and could still fall apart. If completed, the deal would fill NXP's most notable gap: the company already has automotive-grade MCUs, radar processors, and security chips, but lacks the "seeing the road" perception compute — and its automotive business accounts for roughly half of its projected $15.1 billion revenue this year.Source: Startup Fortune source
Luoshi Robotics (2568.HK) | First Interim Profit Alert Since Listing | Revenue No Less Than RMB 400 Million · industrial
A voluntary announcement on July 30 disclosed that H1 2026 revenue was no less than RMB 400 million, up over 127.4% year-on-year, while net loss narrowed to no more than RMB 86 million (down over 4.8% year-on-year); on an adjusted basis, the company swung from a loss of RMB 35.2 million in H1 2025 to a profit. Management attributed the growth primarily to sales of embodied-intelligence robots, projecting the segment would exceed 30% of total reported-period revenue (that segment brought in RMB 47.014 million for full-year 2025). As of June 20, cumulative orders for its embodied-intelligence robots topped 10,000 units, with total confirmed orders across the company exceeding 53,000 units. The company (a Chinese industrial robotics maker) listed on the Hong Kong Stock Exchange on July 9 at an issue price of HK$38, and closed at HK$56.9 on July 30 following the profit alert.Source: 维科网机器人 source
Yotlive | Pre-A+ Round | Investors Revealed · hardware
Following its July 29 disclosure of a completed Pre-A+ round, the round's investors have now been revealed: led by Dinghe Gaoda, with co-investment from listed companies Changshu Automotive Trim and Zulong Entertainment. Yotlive (Chinese smart-textile sensor startup) makes "fabric as sensor" products — weaving metal-drawn alloy yarn into multi-layer fabric to form a sensing matrix, integrally woven in a single process that avoids the inter-layer slippage errors of conventional patch-based sensors. The company previously raised an exclusive angel round from Xiaomi and a Pre-A round led by an Ecovacs-affiliated fund. Its flexible-fabric tactile-sensing data glove is disclosed to launch in August, with several thousand units already in pre-orders.Source: 机器人大讲堂 source
Anhe Chuangyue High-Tech (Suzhou) | Series A+ | Tens of Millions of RMB · hardware
A general machinery-parts manufacturer whose business spans automotive parts, general machined components, and industrial robot manufacturing. The round involved participation from Changshu Guofa Venture Capital, Jinyu Maowu, and others.Source: 证券之星 source
III. Commercialization & Deployment
Pony.ai: robotaxis reach per-vehicle profitability in both Guangzhou and Shenzhen · autonomy ⚠️ Company-disclosed figures
Pony.ai's seventh-generation robotaxi turned a monthly per-vehicle operating profit in Shenzhen this past February, following the same milestone in Guangzhou in November 2025. As of February 28, Shenzhen vehicles averaged daily net income of 338 RMB and 23 daily orders; during the Lunar New Year holiday, average paid orders per vehicle reached 26 per day, versus a nationwide robotaxi average of roughly 15 daily orders in 2025. The profitability came from two levers: the seventh-generation system uses 100% automotive-grade components, cutting materials costs by roughly 70% versus the prior generation, while vehicle utilization and the ratio of remote-assistance staff to vehicles were optimized in parallel. The "per-vehicle profitability" figure here covers per-vehicle operating costs — vehicle and autonomous-driving kit depreciation, charging, maintenance, insurance, ground support, parking, and network fees — excluding company-level expenses like R&D. Shenzhen's operating zone grew from 21.7 square kilometers in March 2025 to 167.4 square kilometers by year-end.Source: 南都·湾财社 source
Aurora lays out its per-mile pricing for driverless trucking: two paths, $2 and $0.85 · autonomy
Aurora posted a Q2 net loss of $270 million on revenue of $2 million (up two-fold year-on-year), with a loss of 14 cents per share, missing analysts' expectation of 12 cents. More telling was CFO David Maday's reaffirmation of two pricing tracks: the all-inclusive Transportation-as-a-Service (TaaS), under which Aurora holds the operating authority, owns the vehicles, and carries insurance, priced at over $2 per mile; and the subscription-based Driver-as-a-Service (DaaS), targeted at over $0.85 per mile, under which customers purchase and manage their own fleets. The company plans to migrate customers from TaaS to DaaS starting in 2027 — CEO Chris Urmson said customers "want to own these assets." Anchoring this transition is cold-chain carrier Hirschbach, which under an April memorandum of understanding will deploy 500 tractors in 2027–2028, roughly one-sixth of its 2,948-vehicle fleet, with final commercial terms expected to be settled this year.Source: FreightWaves source
Atlas and Kodiak scale their driverless heavy-truck fleet from 28 to 100 trucks · autonomy
The most compelling commercial case study for driverless trucking isn't on interstate highways — it's alongside a 42-mile sand-hauling corridor in the Permian Basin. Atlas Energy Solutions announced an expansion of its driverless frac-sand delivery project with Kodiak AI, adding a second loading point and committing to grow the driverless fleet from 28 to 100 trucks by mid-2027. The new loading point means trucks can now depart simultaneously from sites in both Texas and New Mexico (roughly 90 minutes apart), instead of scheduling around a single origin point. As of March 31, the 28-truck fleet was running 15 routes, had completed roughly 7,000 total runs, and had hauled over 450,000 tons of sand — logging over 23,500 driverless hours in Q1 alone; 176 trucks in a single day on July 20 set a record since the deployment began 18 months ago. The two companies have also previously run triple-trailer configurations hauling over 135 tons in total.Source: FreightWaves source
Waymo opens fully driverless rides in Nashville · autonomy
Nashville users are starting to receive invitations for fully driverless rides, initially covering 60 square miles; the airport isn't yet included, with the company saying it will be added "soon." Test vehicles without safety drivers had already been visible locally since this past February, and this marks the formal opening to the public. Rides will initially go through Waymo's own app, with some trips distributed through partner Lyft's app later this year.Source: Mashable source
San Francisco gets $30/hour humanoid home cleaning — with a human sitting in the control room · humanoid ⚠️ Teleoperated, not autonomous
Startup Tau Robotics has launched a humanoid robot house-cleaning service in San Francisco, priced at $30/hour and currently invite-only. Co-founder and CEO Alex Koch says the robots can already handle basic tasks like vacuuming and wiping kitchen counters, with most customers being "people who wouldn't otherwise hire a cleaner." One key caveat: the system isn't autonomous yet — the robot is operated by a human in a central control room, with AI only assisting. Two other details worth noting: the robot's chassis is Unitree's G1, the exact hardware platform the FCC named on July 28 when it added the model to its covered-equipment list (the list doesn't apply retroactively to already-certified units operating in the US, so existing machines can keep running legally); the G1 also has two publicly disclosed, still-unpatched vulnerabilities, CVE-2025-2894 and UniPwn. Separately, the company's terms of service grant it a perpetual, worldwide, royalty-free license to footage the robot captures inside customers' homes, which doesn't terminate even if the account is closed.Source: IT之家 source, Tech Times source
IV. Industry Developments
MIIT: China's industrial robot output hits 538,000 units in H1, up 28.0% year-on-year · industrial
According to H1 electronics-manufacturing operating data released by the Ministry of Industry and Information Technology's Operation Monitoring and Coordination Bureau, China's industrial robot output reached 538,000 units, up 28.0% year-on-year — outpacing growth in new-energy vehicles (6.0%) and integrated circuits (23.1%) over the same period. Value-added output for above-scale electronics manufacturers rose 14.8% year-on-year in the same period, 9.4 percentage points above overall industrial growth; total profits reached RMB 577.7 billion, up 96.9% year-on-year, versus revenue growth of 18.5% — with profit growth far outpacing revenue growth, suggesting this round of improvement is driven more by structural and cost-side factors than by pure volume growth.Source: Ministry of Industry and Information Technology source
Former Amio co-founder and CTO Feng Qian departs to found industrial embodied-AI startup · embodied
Corporate filings show Feng Qian founded Shanghai Xingqianyuan Robotics Technology Co., Ltd. in June of this year, serving as its legal representative and controller, with a similar focus on embodied AI for industrial settings. On July 23, Xingqianyuan signed a strategic cooperation framework agreement and letter of purchase intent with Beijing Ruidekang Technology, covering collaborative development and deployment of a vacuum liquid-crystal-scraper embodied-AI robot for upgrading Ruidekang's production line for its self-developed proton therapy machines (precision oncology radiotherapy equipment). Feng Qian earned his master's and PhD at the Technical University of Munich under roboticist Alois Knoll, and previously worked as an early-stage research scientist at Agile Robots and as a researcher at the German Aerospace Center's Institute of Robotics. Academic researchers striking out to found startups en masse has been a notable pattern this quarter — Amio, the company he left, had itself only just completed its seed-to-angel funding rounds.Source: AI智件 source (WeChat, CN)
Sunday Robotics claims 99.1% success rate for laundry-folding in unfamiliar homes · embodied ⚠️ Vendor-reported figures
Bay Area startup Sunday Robotics has released a new model, ACT-2, claiming its wheeled home robot Memo achieves a 99.1% success rate folding clothes it hasn't seen before, in homes it hasn't operated in — though this figure comes from the company itself, with no third-party replication. Training happens in two stages: humans first collect base skills wearing the company's proprietary sensor gloves, then a fleet of robots iterates from its own failures. The company has also proposed a self-evaluation standard called "Solve," which calls on vendors to disclose testing conditions and the cost of adapting to new environments — though both this standard and the figures above currently come only from vendor self-reporting.
Source: News Anyway
source
Korea's Robros releases backflip video for its IGRIS-C humanoid, timed to the gap left by the FCC ban · humanoid ⚠️ Vendor-reported figures
Robros, a roughly 40-person Seoul startup founded in 2020 that has raised $10.6 million to date, released a video on July 29 showing its bipedal humanoid IGRIS-C performing a windmill spin, backflip, combo kicks, and squats — claiming this is the first bipedal humanoid independently developed by a South Korean company to reach this level of full-body dynamic motion, with hardware, control algorithms, and AI all developed in-house, without foreign actuators, off-the-shelf control frameworks, or external motion capture. The timing is notable: the video came out one day after the FCC added foreign-made humanoid and quadruped robots to its covered-equipment list, and Chinese manufacturers account for roughly 85% of global humanoid shipments. The backflip itself isn't a useful industrial motion — it's a stress test of whether full-body dynamic control can coordinate all joints in real time during an airborne rotation. The company also claims the model is already deployed on a factory production line, though this deployment claim has not yet been independently verified.Source: Tech Times source
Qualcomm explains the Computex humanoid collapse: it was a safety mechanism triggering as designed · humanoid
Following up on NEURA Robotics' 4NE-1 Gen 3.5 humanoid falling backward on stage at Computex, Qualcomm has offered an explanation: the incident stemmed from a brief communication fault, which triggered the robot's built-in "safe collapse" routine — a mechanism designed precisely to minimize the risk of a fall to the surrounding environment and bystanders. The robot runs on Qualcomm's Dragonwing IQ10 robotics reference design, with roughly 700 TOPS of compute. The company says the safety system worked as intended, and the incident data will be used to improve reliability.Source: XimiTime source
Unitree "robot attacks person" video was a staged marketing stunt, agency admits · humanoid
A viral video appearing to show a Unitree robot "going out of control and attacking someone" in China has been verified as a staged performance by a marketing team in Indonesia, which has since admitted the video was scripted content. Once a humanoid robot enters public discourse, false safety scares like this spread cheaply while corrections are costly — this kind of material is a net negative for compliance discussions across the entire industry.Source: 新浪新闻 source
Seedance 2.5 reaches the production line: PDFs become operation videos, flight robots log tens of thousands of simulated flights first · world-model
Following its earlier release and initial round of enterprise integrations, ByteDance's Seedance 2.5 now has more concrete industrial use cases: a single generated video can run up to 30 seconds and accept up to 50 multimodal inputs combined. XCMG (Chinese construction-machinery maker) converts product materials, SOPs, and repair manuals directly into standardized operation videos, cutting the cost of reshooting whenever a process changes. Micro Multicopter Aero Technology (Chinese industrial drone maker) uses synthetic video to fill in long-tail flight scenarios, while LingChu Intelligence (Chinese embodied-AI startup) adjusts lighting, background, and other elements to generate robot training data. The 30-second length is significant because it's just long enough to cover a complete action sequence in embodied-AI settings, and long enough to capture a complex intersection crossing in autonomous-driving settings.Source: 搜狐 source
Hardware & Supply Chain
· Inovance Technology (Chinese industrial automation maker): at a July 28 institutional research briefing, disclosed that it has developed all six core humanoid components in-house — frameless torque motors, low-voltage high-power drivers, planetary roller screws, rotary actuators, linear actuators, and a seven-DOF bionic robotic arm, with the roller screws reaching C3-grade precision. Several products have already secured customer design wins, and the seven-DOF arm is in joint testing and moving toward small-batch shipments; the company says it won't build complete robots itself source (WeChat, CN)
· Tacta Systems: the Palo Alto dexterous-hand company unveiled its TactaBot platform, whose hand uses a proprietary "fluidic tendon" drive with 15 independently actuated joints and force/temperature tactile sensing, targeting durability over millions of manufacturing cycles; its companion Tacta Glove captures workers' force, motion, temperature, and video data during tasks, with deployment to manufacturing customers planned for early 2027; the company has raised $75 million to date (from SoftBank, Matter Venture Partners, and others) source (WeChat, CN)
· AMD: unveiled its Kria AI system-on-module and robotics development platform at Advancing AI on July 23, built on Ryzen AI Embedded X100, unifying CPU/GPU/NPU/FPGA under a shared memory architecture; benchmarked against Nvidia's Jetson Thor T5000, AMD claims 3.4× faster real-time processing and 2.3× more concurrent agents (⚠️ vendor-reported figures); the software stack centers on ROS 2 and ROCm, and Foundation Future Industries' Phantom humanoid has already migrated to it from an Nvidia/Intel-based setup source
· Topband (Chinese motor and component maker): disclosed that its supply relationship with UBTECH centers on hollow-cup motors, as well as integrated "reducer + encoder + lead screw" actuator assemblies source
Top comments (0)