Today's Highlights
· NeoteAI (Chinese embodied-AI startup) and Fudan University jointly open-source 30,000 hours of visuo-tactile data, releasing a tactile foundation model plus VTLA and TWAM in one batch
· Fei-Fei Li, Abbeel and others co-author T-Rex dexterous-hand model, hitting 65% success vs. 35% for vision-only baselines
· NHTSA proposes scrapping the mandatory brake-pedal rule, potentially ending the 2,500-vehicle/year capacity ceiling on robotaxis
· Qianxun Intelligent (Chinese embodied-AI startup) closes a RMB 1.5 billion Series A+, bringing its total across four rounds in three months to over RMB 4.5 billion
· Uber Eats partners with Serve to roll out up to 2,000 sidewalk delivery robots across the US market
I. Research Progress
30,000 hours of visuo-tactile data: elevating touch from "auxiliary modality" to foundation-model status · manipulation
Scaling tactile sensing has long been stuck on the same bottleneck: gel-deformation images, capacitive matrices, and six-axis force vectors each speak their own "dialect," so data from different devices can't be folded into a single model. NeoteAI (Chinese embodied-AI startup) and Fudan University's Institute of Trustworthy Embodied AI answer this head-on with a shared data foundation plus two model tracks — NeoData comprises over 30,000 hours of visual-tactile interaction video, roughly 1.4 million manipulation clips and 3.3 billion timesteps, spanning 6 robot embodiments including Franka, Piper, and UR5e across 450 real-world long-horizon tasks, with the first 5,000 hours already open-sourced; the companion NeoForce unified-representation model normalizes signals across sensor types. VTLA's approach ditches lagging current-tactile readings in favor of predicting tactile evolution 50 steps into the future — achieving an 85% success rate on plugging in a connector versus 60% for comparable vision-only approaches, and 99% vs. 35% on key removal. TWAM instead folds touch into a world model, using three asynchronous experts for video, touch, and action to jointly predict all three sequences; at 7.2 billion parameters it reaches an average simulated success rate of 84.5%, against a strongest world-model baseline the paper cites at 36%.
NeoteAI × Fudan University Institute of Trustworthy Embodied AI · Three technical reports released together · Analysis: JiqiZhixin (Synced) source (WeChat, CN)
T-Rex: giving touch its own high-speed lane, doubling dexterous-hand success rates · manipulation
The author list rarely brings together Fei-Fei Li, Trevor Darrell, Jitendra Malik, Pieter Abbeel, Ken Goldberg, and NVIDIA GEAR Lab's Jim Fan — yet the focus here isn't a general-purpose brain but the dexterous hand. The paper opens with a counterintuitive finding: stitching tactile force signals directly into a pretrained VLA actually drops success rate from 17% to 6% — touch can't just be bolted on. T-Rex's fix is a Mixture-of-Transformer-Experts design: a 1.41B-parameter Action Expert handles low-frequency motion planning, a 0.62B-parameter Tactile Expert handles high-frequency refinement, and the 10-step diffusion denoising process is split at step 4, with real-time tactile signals injected only in the later steps; touch itself is first compressed via VQ-VAE, turning 16 frames of past force history into 64 discrete codewords that filter out sensor zero-drift. Across 12 real-world dexterous manipulation tasks, average success rate reaches 65% versus 35% for the strongest vision-only baseline, with lock-picking at 70% vs. 0%; removing the tactile input drops performance back to 42%.
Stanford / Berkeley BAIR / NVIDIA GEAR et al. · Analysis: QbitAI source (WeChat, CN)
Riemann-1.0: a world-action model spanning 232,000 hours and 21 embodiments · world-model
Riemann Dynamics, a unit of Kunlun Tech, has released Riemann-1.0, unifying first-person human video, UMI teleoperation demonstrations, and heterogeneous robot trajectories into 232,000 hours of training data spanning 21 embodiment types, built on a fully causal world-action modeling architecture with three-stage progressive embodied pretraining. The team reports over 99% success on LIBERO, an 8-percentage-point improvement over prior methods on RoboCasa's 365 household scenarios, and a 60% gain in long-horizon generalization; it also claims that adding large volumes of heterogeneous data noticeably improved even tasks like folding clothes, previously thought to require real-robot data exclusively. ⚠️ Vendor claim
Kunlun Tech · Riemann Dynamics · Analysis: KejiQianyanquaner source (WeChat, CN)
MoWorld: a real-time 4D world model running end-to-end on Chinese NPUs · world-model
Mainstream interactive world models typically max out at 5–10 FPS for real-time inference and are tied to high-end NVIDIA GPUs. MoWorld, a collaboration between Moxin Technology (Chinese chip startup) and the team of Academician Pan Yunhe at Zhejiang University, moves the entire pipeline — training, distillation, and edge deployment — onto Chinese-made NPUs; its 14B MoE model reaches up to 50 FPS, at 30%–50% of the overall inference cost of a traditional GPU setup. The key techniques are curriculum-based cross-frame pretraining that progressively extends generation length from 125 frames to 2,000 frames, plus unified-latent-space distillation that compresses 50 denoising steps down to 4 and rewrites the bidirectional DiT into a self-driven causal DiT.
Moxin Technology × Zhejiang University · arXiv 2607.06216 source · Analysis: AIGC Studio source (WeChat, CN)
Two humanoid robots complete a remote surgery for the first time · manipulation
Engineering and surgical teams at UC San Diego completed two surgeries in a preclinical study using teleoperated humanoid robots: in one, a humanoid robot led the operation with a human surgeon assisting to remove a gallbladder; in the other, two humanoid robots worked together to perform the same procedure on a large non-human primate. The robot, code-named Surgie, stands about 5 feet tall and weighs 60 pounds, whereas existing surgical robot systems typically weigh around 1,800 pounds and require a dedicated operating room plus a full pre-surgical setup team — that gap is exactly what the researchers are betting on: humanoid robots work in standard operating rooms with standard instruments, potentially enabling use in remote communities or search-and-rescue settings that can't accommodate traditional equipment.
UC San Diego (Michael Yip et al.) · Preclinical study · Reported by: Earth.com source
HEAR: adding a "hearing" pathway for sound-centric manipulation tasks · manipulation
Nearly all existing VLA perception and control mechanisms are designed around vision, where scenes are temporally continuous and can tolerate some system latency; HEAR instead targets manipulation tasks whose success criteria mainly hinge on sound, proposing a vision-sound-language-action paradigm.
IRMV Lab · Accepted to IJRR 2026 · Analysis: IRMVLab source (WeChat, CN)
Act2Answer: VLAs haven't lost their common sense — it just isn't reaching their hands · benchmark
Testing a VLA's world knowledge with a task as simple as "place the block on the correct answer" produces a counterintuitive result: after fine-tuning a VLM into a robot policy, its knowledge is not forgotten — it remains latent in the intermediate-layer representations — but it can't be transferred to the action end for actual use. The evaluation question thus shifts from "does it still remember" to "can it actually use it".
Analysis: XiaohuXueLLM source (WeChat, CN)
Other papers today: LingBot-VLA 2.0, an open-source VLA update from Ant Group's Lingbo team focused on cross-task and cross-embodiment generalization without retraining when switching robot bodies; Anchor-Align, which finds that standard behavior-cloning fine-tuning overwrites a VLM's visual and semantic priors, and instead uses anchored alignment to preserve out-of-distribution generalization.
Open Source · Tools · Benchmarks
· NeoData / NeoForce: 30,000-hour visuo-tactile interaction dataset and cross-sensor unified representation model, with the first 5,000 hours already released source (WeChat, CN)
· DreamX-World 1.0: a 5B-parameter interactive world model from Amap's DreamX team, with code and weights fully open-sourced, supporting camera navigation, region revisiting, and promptable events; generates at 16 FPS streaming on 8 RTX 5090 GPUs source (WeChat, CN)
· A JAX/Flax reimplementation of Dreamer 4: a third party has filled in and open-sourced training recipe details that DeepMind hadn't fully disclosed, lowering the bar for reproducing the world model source (WeChat, CN)
II. Funding & Deals
Qianxun Intelligent | Series A+ | RMB 1.5 billion | Over RMB 4.5 billion across three months · embodied
This round was joined by top-tier USD funds and large industrial investors, with investors close to the deal describing the allocation as "highly sought-after." This is Qianxun's fourth round in three months — in February it raised nearly RMB 2 billion across two rounds, pushing its valuation past RMB 10 billion, then a month later raised another RMB 1 billion at a valuation above RMB 20 billion (co-led by Shunwei Capital and Yunfeng Capital, marking the first time Lei Jun and Jack Ma have both made major investments in the same embodied-AI company), bringing the cumulative total to over RMB 4.5 billion as of this round. The company simultaneously claims its self-developed Spirit v1.6 has topped the RoboArena leaderboard — jointly launched by Berkeley, Stanford, and NVIDIA — outranking Cosmos 3 and Pi0.5. Proceeds are earmarked for iterating its embodied foundation model, building out global real-world data infrastructure, and scaling deployment across multiple industries; its data-collection system currently uses nearly 1,000 wearable devices collecting simultaneously across more than 100 cities. Source: 36Kr source
Motion Brain | Additional Pre-A funding | Nearly RMB 100 million | Valuation up over 10x in six months · world-model
Investors include industrial investment platform Jinyue Investment, Chuanghehui Capital, a leading property-services company, and a Hong Kong consortium, with existing shareholder Xuhui Capital following on. This is a second injection within two months, following a RMB 300 million Pre-A round in May; another nearly RMB 500 million Pre-A+ round is in the process of closing. The company was co-founded by Fudan University Deep Learning Lab director Chen Tao, former Intel China Chief Scientist Zhang Yimin, and serial entrepreneur Mu Zelin, and pursues a "world-action model" approach distinct from the mainstream VLA route, with a training recipe of 80% internet video + 10% motion capture + 10% real-robot data — the company says this cuts real-robot data requirements by 90%. On the edge-deployment side, it has reduced inference latency from roughly 200 milliseconds to 10 milliseconds, and per-unit inference cost from RMB 200,000 to RMB 10,000. It is also one of the few China-based embodied-brain companies with auditable revenue already — RMB 10 million collected in 2025, and RMB 30 million in H1 2026. Source: Yingke (36Kr) source (WeChat, CN)
Magnendo (US) | Series A | USD 18 million (approx. RMB 120 million) · adjacent
Led by thrombectomy-device maker Penumbra, with several new and existing institutions participating. Magnendo is an MIT spinout, with technology originating from Xuanhe Zhao's group's work on programmable magnetic-domain soft materials and ferromagnetic soft continuum robots — the team first proposed printing programmable magnetic-domain soft materials in 2018, earning a Nature cover story. The company is led by CEO Yoonho Kim, with Zhao serving as scientific advisor. Proceeds will fund productizing its magnetic robotic navigation platform and preparing for first-in-human clinical studies, targeting the most time-consuming step in neurointerventional thrombectomy — steering a guidewire through highly tortuous blood vessels to reach the clot.Source: Keyan Lidao source (WeChat, CN)
HaloBraid (US) | New funding round | USD 7 million · adjacent
A hair-braiding robotic assistant led by Nigerian engineer Yinka Ogunbiyi, aimed at mechanizing the highly repetitive, time-consuming manual work of salon hair braiding.Source: Head Topics source
Industry data: dexterous-hand funding in China surpasses RMB 25 billion in H1 · hardware
According to incomplete statistics from IT Juzi, total funding raised by China's dexterous-hand sector in all of 2025 was RMB 16.877 billion, while H1 2026 alone has already surpassed RMB 25 billion. The industry backdrop: dexterous hands account for 18%–25% of a complete humanoid robot's cost, making them the single priciest component and a direct driver of overall unit pricing.Source: AING Yingji source (WeChat, CN)
III. Commercialization & Deployment
Uber Eats partners with Serve Robotics, deploying up to 2,000 delivery robots · adjacent
The partnership expands from a pilot in West Hollywood a year ago to multiple US markets, with a scale cap of 2,000 units. Backing the expansion is real pilot-phase traction: robot-fulfilled Uber delivery orders through Serve grew over 30% month-over-month, with more than 200 restaurants onboarded in Los Angeles alone. Serve's sidewalk robots claim L4 autonomy, capable of self-rerouting around obstacles; the company has previously piloted with Walmart (Arkansas) and Pizza Hut (Vancouver), and has been delivering for 7-Eleven in Los Angeles since 2021. Compared with earlier scattered pilots, this deal moves delivery robots into a signed, at-scale rollout phase.Source: Mashable source
Zhiyuan Robotics (Chinese humanoid startup) unveils Genie G2 Max with JD Logistics, 18kg single-arm payload entering "Smart Wolf Warehouse" · humanoid
The two companies describe this as the world's first safety-certified heavy-payload humanoid robot, with an 18kg single-arm payload, first deployed at JD's "Smart Wolf Warehouse." The payload figure is the real news here — most humanoid robots' single-arm load capacity still sits in the single-digit to low-teens-kilogram range, and 18kg is enough to handle genuinely heavy-lift categories in warehouse operations.⚠️ Vendor claimSource: Zhanlue Gongfang source (WeChat, CN)
On produces a shoe upper in 3 minutes with robots, cutting 200–300 hours of manual labor · industrial
At On's manufacturing lab at its headquarters in Zurich, Switzerland, a robotic arm rotates a foot-shaped mold while a nozzle continuously extrudes 1.5 kilometers of molten plastic filament, winding it layer by layer into a running-shoe upper — the whole process takes about 3 minutes. Athletic shoes have long relied on cutting, sewing, and gluing to assemble as many as 200 separate parts; On's LightSpray process compresses a shoe's construction down to just 7 components, and the company estimates this saves 200–300 hours of labor across the entire development and production cycle. Co-CEO Caspar Coppetti emphasized this isn't simply bolting robotics onto an old shoe design: "We designed a process that only a robot could make" — which is also what distinguishes it from most production-line automation retrofits: the robot came first, and the product form followed.Source: Fortune source
An open-pit coal mine in Inner Mongolia runs 100 driverless electric mining trucks, discloses a year of operating data · autonomy
The 100 XCMG ZNK95 driverless all-electric mining trucks (branded "Huaneng Ruichi") at Huaneng Yimin's open-pit coal mine have been in operation since May 2025, running routinely without a driver in the cab, centrally dispatched from a single control room over a dedicated Huawei 5G-A network with roughly 500Mbps uplink and about 20ms latency. Recently disclosed operating data includes: battery swaps completed in under 6 minutes, and operation down to temperatures as low as -48.5°C. Huaneng states the fleet has set records for load capacity, speed, and cold tolerance, with operating efficiency roughly 20% higher than human-operated crews.⚠️ Single-party claimSource: Interesting Engineering source
XPeng's Guangzhou factory production line enters final joint testing · humanoid
Following the earlier start of small-batch trial production of humanoid robots at its Guangzhou factory (previously reported), XPeng has confirmed to China Business Journal that its mass-production line has entered final joint-debugging stage.Source: China Business Journal source (WeChat, CN)
IV. Industry Developments
NHTSA proposes scrapping the mandatory brake-pedal rule, potentially removing the robotaxi capacity ceiling · autonomy
Federal Motor Vehicle Safety Standard FMVSS No. 135 has required, since 1995, that every light vehicle's service brake be actuated by a foot pedal — a rule that now blocks vehicles like Tesla's Cybercab and Amazon's Zoox, which were designed from the outset without a steering wheel or pedals. NHTSA's proposed rule change (docket number NHTSA-2026-0728) would remove the manual-brake-control requirement for vehicles built solely to operate under an automated driving system; the public comment period closes at the end of business on July 27, with about 52 comments on the docket as of this report. Braking-distance performance requirements remain unchanged, only the measurement starting point shifts to when the electronic command signal is sent to the brake control. The real weight of this change lies in the existing rules it replaces: pedal-less robotaxis currently can only operate under a temporary exemption via 49 CFR Part 555, capped at 2,500 vehicles per manufacturer per year — workable for demonstration fleets, but structurally incompatible with mass commercial deployment. If finalized as proposed, this would be the first permanent federal approval pathway for robotaxis without pedals or a steering wheel.Source: Tech Times source
KAIST and NVIDIA co-establish a physical-AI center in South Korea, starting with a human-motion foundation model · adjacent
The two organizations will establish a "Human Physical Intelligence Technology Center" within KAIST's Department of Mechanical Engineering, jointly designated an NVIDIA AI Technology Center (NVAITC), co-directed by wearable-robotics researcher Kyungchul Kong and bio-robotics researcher Jeong Kim. The first project is to train a human-motion foundation model on large-scale human motion data to understand, predict, and generate human movement — the goal being to let wearable and humanoid robots accurately recognize a user's intent and movements and respond naturally, with applications in rehabilitation and daily assistance for people with mobility impairments, motor-function analysis, and digital healthcare. The collaboration also includes research and teaching on the Omniverse and digital-twin platforms, a student ambassador program of 5 to 10 students, and biannual reviews of research goals and outcomes plus an annual international symposium. This aligns with Hyundai Motor Group's direction this week: physical AI is moving from corporate slogan to institutional-level infrastructure investment in South Korea.Source: Digital Today source
Nikkei xTECH tears down Unitree's G1: "won't catch up anytime soon" · humanoid
On July 23, Japanese tech outlet Nikkei xTECH published a complete teardown video of the Unitree G1, with engineers dissecting it joint module by joint module, from wiring harness layout to lightweight structural design, trying to understand why Chinese robots are progressing so fast. What comes up repeatedly in the video is the machining and integration quality — the smooth feel of the screws, the miniaturization of the battery while maintaining high output power — and the conclusion is that closing the gap in the short term for Japan is "probably unrealistic." The story was reposted by dozens of accounts on Chinese platforms today, drawing attention far beyond its actual information content; placed back in industry context, the real variable is localization rate and cost structure, not the teardown engineers' impressions.Source: EE Times China source
New York school-district "robot teacher" update: manufacturer revealed to have adult-doll business ties · adjacent
Following the earlier state education department pressure letter and the district's pause on the rollout (previously reported), a new disclosure fills in a key detail: the robot, code-named Sally and priced at USD 57,000, comes from Realbotix, whose parent brand RealDoll has long made adult-doll products, and whose USD 175,000 "companion robot" Aria has already drawn controversy. The Salamanca school district in western New York had planned to have Sally serve as a teaching assistant for high-school STEM and arts classes, providing real-time translation in over 100 languages, generating course materials, and offering prompts when a teacher's lesson stalls. Superintendent Mark Beehler's rationale: "Students can always find a way around the rules a school sets — what the school should teach is proper use, not an outright ban."Source: Futurism source
Tesla robotaxi drives the wrong way down a one-way street in Tampa · autonomy
The day after the service launched in Tampa, a bystander-submitted video captured a Tesla robotaxi with no one inside stopped and driving the wrong way on the one-way North Armenia Avenue. The witnessing driver said they initially assumed the driver didn't know the road, only to find the driver's seat empty after passing. Tesla had not responded as of this report.Source: AOL/WFLA source
BYD sets early-August date for humanoid robot reveal · humanoid
Following its earlier announcement of entering the humanoid space with retail-store guidance as the first application scenario (previously reported), BYD's Zhengzhou "Di Space" has released a teaser poster stating its humanoid robot product will make its first public appearance in early August, standing 1.61 meters tall and able to speak 6 regional Chinese dialects.⚠️ Vendor claimSource: Sohu source
Hyundai Motor establishes an AVP unit in Silicon Valley · adjacent
Following its announced transformation into a "Physical AI Solutions company" (previously reported), Hyundai Motor Group has established an AVP unit in Silicon Valley, deepening ties with the San Francisco innovation ecosystem and moving physical-AI research and talent forward to the US West Coast.Source: Seoul Economic Daily source
Hardware · Supply Chain
· Tactile-sensor price anchor: PaXini Tech (Chinese tactile-sensing company) has brought the price of a single tactile sensor down from the six-figure-RMB range of early imported units to as low as RMB 199, closed a Series B of over RMB 1 billion in March at a post-money valuation above RMB 10 billion, and is preparing a Hong Kong IPO — post-listing financial disclosures will for the first time provide third-party-audited real shipment figures source
· Tashan Technology: closed a Series B of several hundred million RMB in July (its third round in six months), with industrial players including Joyson Electronic, AUX, and Pengling shares joining; monthly tactile-sensor deliveries have reached tens of thousands of units, with H1 order volume already over four times all of last year — the company is NVIDIA Isaac Sim's first global tactile-simulation partner source
· Huaweike: China's top domestic dexterous-hand tactile-sensor shipper by volume with over 70% market share, minimum trigger force of 0.05N; its planned tens-of-millions-of-units-scale robot e-skin production base in Gedian, Ezhou (Hubei) is expected to bring its first production lines online in October. Nengsida (a Hanwei Technology subsidiary) has expanded capacity to 20 million flexible tactile sensors per year with yield stabilizing above 92%, and passed IATF 16949 automotive-grade certification in February source
· Actively-cooled joint modules: Shiyue Technology has launched a miniature joint module with active cooling, achieved by redesigning the motor configuration to create airflow channels, claiming a temperature reduction of over 50% and extending no-load continuous-use lifespan to over 5,000 hours — directly addressing the short-lifespan issue previously flagged for dexterous hands source (WeChat, CN)
Top comments (0)