Today's Highlights
· Pony.ai's Q2 Robotaxi revenue up 691% YoY, fleet expands to 1,975 vehicles
· Tesla Cybercab could open to riders in Austin as soon as this month, starting with employees
· Apollo Go covers 28 cities, fully driverless mileage exceeds 240 million km
· LG partners with Nvidia to amass 100,000 hours of humanoid training data by year-end
· GigaBrain-0.7 pretrained on 37,000 hours of embodied data, weights to be open-sourced
· Sany's electric autonomous mining trucks make first export to South America, 300+ units deployed cumulatively
· Integral AI, founded by ex-Google researchers, folds after raising only about $5 million in seed funding
1. Paper Progress
GigaBrain-0.7: A three-system architecture pushes embodied foundation models toward "cross-embodiment generality" · world-model
Today's must-read: it advances two open questions at once — how the architecture should be designed, and how large the data can scale. GigaBrain-0.7 unifies understanding, prediction, and action into three systems, scales pretraining to 37,000 hours of heterogeneous embodied data, and introduces single-stage alignment training so vision-language understanding and multi-embodiment action generation are optimized jointly. The authors report it surpasses the earlier GigaBrain-0 series and π0.5 on zero-shot base capability, language-instruction following, and post-training task success rate, covering both home and industrial scenarios on its own Maker H01 platform as well as mainstream embodiments. All training code and pretrained weights will be open-sourced — a factor that will determine whether it can actually be reproduced.
GigaBrain Team · arXiv 2608.15875 source
τ0-VLA: letting robots "think a bit longer" on the hard step · vla
High-level decision-making in hierarchical VLA models typically runs a single forward pass, spending the same compute on hard and easy problems alike. τ0-VLA reframes subtask generation as a scalable reasoning problem: the high-level policy uses a world model to search over several candidates at critical junctures before committing, while the low-level policy executes across multiple embodiments. Trained jointly across modalities on 40,115 hours of real heterogeneous data, experiments show that adding test-time compute noticeably improves next-subtask prediction accuracy, translating into higher closed-loop success rates on long-horizon tasks.
Xiaowei Cai et al. · arXiv 2608.16885 source
Zetta ζ: freeze the base policy, let the "execution shell" evolve on its own · vla
Most existing embodied-agent frameworks are open-loop — running a fixed skill sequence to completion before reviewing, unable to manage real-time state during execution. Zetta keeps the base policy frozen and evolves a codified runtime critic and recovery skills online, using three time-scale-separated loops that handle action-frequency-level control, rollout-level critique and recovery proposals, and skill updates gated by verification, respectively. It achieves 90.8% and 93.6% success rates on LIBERO-Pro and RoboCasa respectively, with inference sped up 11.1x; success rate continues to rise with autonomous exploration experience, and learned skills transfer zero-shot.
Xin Ding et al. · arXiv 2608.16590 source
Flip a few bits and a VLA drops to zero: the first bit-flip attack targeting embodied models · vla
This paper formally elevates "weight integrity" to a security boundary for embodied foundation models. The authors demonstrate the first bit-flip attack against VLAs: under Rowhammer-style faults, flipping just a few gradient-selected INT8 bits can drive closed-loop success rate to 0%, while hundreds of random flips are nearly harmless. Vulnerability is determined by the action-head architecture — direct regression heads and token-based policies need only 1–5 flips, while flow-matching policies require roughly 100–300; the authors' proposed targeted escape loss compresses π0's budget from about 1,000 flips to about 100. In real-robot validation, K=100 flips caused all 20 trials to fail, versus 14/20 for the clean model. On the defense side, protecting just 3.1% of weights preserves 60% success rate at K=100.
Yudong Gao et al. · arXiv 2608.15475 source
SPD: pretraining data for dexterous hands, collected entirely in simulation · manipulation
Multi-fingered dexterous hands have long been data-starved — real-robot teleoperation is expensive, while human-hand videos are cross-embodiment and require lossy pose estimation and retargeting. SPD lets operators directly manipulate virtual objects via VR headset, collecting same-embodiment trajectories without using any real robot time: 5 operators gathered 75 hours of multi-task data in one week, used to pretrain a causal transformer. On a 56-degree-of-freedom bimanual dexterous-hand platform, fine-tuning with just 1–2 hours of real-robot demonstrations outperforms behavior-cloning policies trained from scratch.
Sarthak Kamat et al. (Chelsea Finn's group) · arXiv 2608.15917 source
CoRe: without failure data or retraining, a VLA "thinks" its way back on track at inference time · vla
Existing recovery methods mostly require failure data, policy retraining, or an external correction agent, all at considerable cost and risk. CoRe is a training-free framework: upon detecting a deviation, it imagines — via synthesized observations — what it would look like for the policy to continue from a recent feasible state, then makes a minimal-magnitude alignment of the robot and scene to that imagined trajectory before handing control back, with no physical trial-and-error involved. Across multiple simulators, VLA backbones, and real-robot setups, it improves success rate by up to 85.0 percentage points and reduces physical resets by 42.2%.
Yanyan Zhang et al. · arXiv 2608.14822 source
EgoTac: "seeing" touch from first-person video · perception
A growing amount of first-person human data used for robot learning lacks a tactile channel, while tactile data itself remains hard to collect at scale. Trained on 5.7 million image-tactile pairs, EgoTac predicts continuous force values and binary contact directly from first-person video, achieving in-domain average force error below 0.06N and out-of-domain contact prediction that outperforms the previous best contact estimator, with zero-shot prediction on unconstrained real-world video. Scaling analysis shows both data diversity and volume are still delivering steady gains.
Wenkang Zhang et al. · arXiv 2608.15060 source
RoboStriker: turning humanoid boxing into a zero-sum game in latent space · locomotion
When multi-agent reinforcement learning acts directly on raw joint space, humanoid robots tend to lose balance catastrophically before developing any real tactics. RoboStriker first distills the tracking capability of preset boxing motions into a topologically constrained latent manifold, then performs neural fictitious self-play within this structured space, constraining policy exploration to the decoder-reachable action manifold. In experiments it substantially reduces the catastrophic imbalance seen in raw-action-space methods, achieves higher win rate and striking efficiency, and has been deployed and validated on a real humanoid robot.
Kangning Yin et al. · arXiv 2608.16195 source
HiPHI: a 600-hour, sub-millimeter-precision dataset of full-body motion and object interaction · benchmark
Internet video lacks precise physical state, while lab motion-capture data is high-fidelity but narrow in behavioral coverage — this mismatch has long constrained scalable learning of humanoid policies. HiPHI uses optical motion capture to collect 600+ hours of full-body motion data, with marker tracking at sub-millimeter precision and object trajectories at mesh-level fidelity, organizing human motion primitives via the linguistic framework FrameNet to systematically maximize coverage. An accompanying evaluation suite measures action-space diversity, interaction grounding, and object consistency, among other dimensions; the authors report it substantially expands motion coverage while maintaining high-fidelity interaction quality.
Jiahao Ji et al. · arXiv 2608.16222 source
Other papers today: ReForce retargets the contact forces from human-hand demonstrations along with the motions themselves onto robot actions, improving multi-finger contact quality (arXiv 2608.15560 source); HAF adapts general-purpose VLAs to humanoid full-body loco-manipulation using hierarchical action flow plus spectral-latent-space RL, outperforming single-stage baselines on 7 real-robot tasks (arXiv 2608.16837 source); a paper re-examining action chunking finds that long open-loop execution mainly helps short-context policies imitate "non-Markovian demonstrations," while with sufficiently long context the most reactive closed-loop policy performs best (arXiv 2608.15938 source); a survey on vision-based tactile sensing (VBTS) treats hardware, learning, simulation, and datasets as one integrated system (arXiv 2608.15490 source); MatchingPolicy decouples demonstration-scene matching from policy learning, improving cross-object contextual imitation (arXiv 2608.16715 source); a study learns urban navigation policies from web-scale in-the-wild first-person video and systematically exposes its long-tail failure modes (arXiv 2608.16476 source); DeepInsight II extends the evaluation pipeline from benchmark execution all the way to matched real-robot evidence and repairability diagnostics (arXiv 2608.16556 source); a 29-DoF Unitree G1 fitted with a three-fingered dexterous hand throws a spin-stable American football spiral, achieving 93.6% spin efficiency (arXiv 2608.16642 source); also a survey on the safety attack surface of foundation-model-driven embodied agents (arXiv 2608.16843 source) and on neuro-symbolic embodied agents (arXiv 2608.16794 source).
Open Source · Tools · Benchmarks
· 51WORLD AperData / AperOne: Beijing-based 51WORLD (Chinese digital-twin/simulation firm) released an embodied-data foundation and application platform — the AperEgo head-mounted capture device plus AperOs integrated hardware-software stack — priced at 5,100 yuan per set at launch, with mass production delivery starting in September; the company claims collection efficiency over 10x that of real-robot teleoperation and 99% data delivery precision (vendor claim), and has also formed a data-infrastructure joint venture with United Imaging source
· Foxglove agentic data platform: launched an agentic data platform for Physical AI at Actuate 26, partnering with NVIDIA on semantic search source
· Trossen Robotics × Stereolabs: Ouster subsidiary Stereolabs' high-fidelity stereo vision cameras integrated into Trossen's AI platform for embodied data collection source
2. Funding and Deals
FORT Robotics | SPAC merger listing | pro forma enterprise value $556.6M | pre-money equity $500M · adjacent
FORT Robotics signed a definitive merger agreement with Newbury Street II Acquisition Corp (Nasdaq: NTWO); the combined company will be renamed FORT Robotics Holdings and trade under the new ticker FROB on Nasdaq. Founded in 2018, the company builds a functional-safety layer for robots — certified to IEC 61508 SIL 3, holding 25 patents, with 600+ customers including Agility Robotics, Google DeepMind, Zoox, Ocado, and DoorDash; investors include Tiger Global and Mark Cuban Companies. In May the company acquired teleoperation firm Mapless AI to add human-in-the-loop remote control. That safety and compliance — a "non-performance" layer — can support a half-billion-dollar valuation signals the industry is starting to pay for the preconditions of scaled deployment. Source: PR Newswire source
LimX Dynamics | planned Hong Kong IPO | up to $300M · humanoid ⚠️ unconfirmed report
Following a Pre-IPO round of nearly $200 million in late July that valued it at RMB 15 billion post-money, the Alibaba-backed humanoid and legged-robot company is reportedly planning a Hong Kong listing, raising up to $300 million. Source: Bloomberg source
Velaura AI (US) | Series A | $110M · hardware
Velaura AI closed a $110 million Series A to advance next-generation ultra-low-power AI compute infrastructure. The power budget for on-device robot inference is far tighter than for data centers, and companies specializing in low-power compute are being repriced by embodied and edge use cases. Source: Business Wire source
Acorn Robot | angel round | investors disclosed | two rounds total in the hundreds of millions of yuan · embodied ⚠️ company claim
Following disclosure of an angel round earlier this month, the investors have now been named as jointly led by China Merchants Capital and NIO Capital, with follow-on from the Tsinghua Shuimu Alumni Seed Fund — just four months after a roughly RMB 100 million seed round in March, bringing the two rounds' total to the hundreds of millions of yuan. The same day, the company unveiled the "Natus AGE-0 embodied instinct model," which it describes as the world's first general-purpose manipulation foundation model centered on tactile perception that does not rely on massive labeled-data pretraining, aimed at zero-shot generalization across embodiments and materials — a route that diverges directly from the mainstream VLA data-scaling paradigm, though its actual performance still awaits third-party verification. Commercially, it has completed a proof-of-concept on a production line for a leading global cosmetics ODM, taking about two months from contract signing to validation. Source: Gasgoo source
CyboPal (Jiangsu Cell Wall Intelligence) | new round | hundreds of millions of yuan · embodied
Led by Oriental Fortune Capital, with follow-on from Legend Star and Tenqi Capital, funding will go toward mass production and delivery of its first product and iteration of its embodied interaction system; the company closed a tens-of-millions-yuan angel round just last July. Its product, CyboPal ONE, is a 27-inch 4K display mounted on a 6-DoF robotic arm that can actively adjust to follow a user's posture within 4 seconds, supporting gesture and voice interaction; it has begun presale on Kickstarter at an early-bird price of $1,499. Founder Peng Tianfang graduated from Tsinghua's Department of Precision Instruments; the engineering team numbers over 100. Source: iFeng Tech (智东西) source
DIGIT Robotics | Pre-A strategic round | RMB 100M · humanoid
Led by a consortium of industrial investors, with existing shareholder Chinaequity Group increasing its stake, and multiple existing shareholders indicating they will continue to follow on in future rounds. Funds will go toward mass production of its bionic humanoid series and ecosystem-building for its scene-brain system RoboEase. Founded in March 2024, the company focuses on high-precision facial-expression control and a fast-slow "dual-brain" architecture; its bionic humanoid Xialan R03, unveiled at WAIC 2026, has over 40 degrees of freedom in the face and can perform 30-plus complex expressions, supporting wake-word-free conversation. RoboEase uses graphical, no-code process orchestration and has been deployed in banks, elder-care facilities, government service halls, and hotel front desks. Source: Gasgoo source
Jiaotong Robot | strategic funding | tens of millions of yuan · industrial
Led by listed company Hoto Tech (002402.SZ), with follow-on from Daohe Fund (whose LPs include Dongfang Precision and Leju Robot) and Jinmu Capital. The company makes industrial embodied-AI mobile chassis, having served nearly 100 Fortune Global 500 companies cumulatively and delivered over 2,000 projects, accumulating 270 million hours of industrial field operational data. Its core omnidirectional mobile chassis can move under unified VLA/VLM command, currently delivering 150 to 200 units a month, and is working with Leju Robot on an industrial proof-of-concept for a dual-arm, 15kg-payload humanoid robot. Source: Ruicaijing source
Unichem acquires Loomia | terms undisclosed · hardware
Auto supplier Unichem acquired flexible-electronics company Loomia to accelerate its entry into the humanoid robot "electronic skin" market. Loomia was previously known for smart-skin development kits aimed at humanoid robots, working on the layer that puts tactile sensing across a robot's body surface. Source: The Robot Report source
3. Commercialization and Deployment
Pony.ai Q2: Robotaxi revenue up 691% YoY, fleet at 1,975 vehicles · autonomy
Pony.ai's total Q2 revenue reached $36.2 million, up 68.8% YoY; of that, Robotaxi service revenue was $12.1 million, up 691.2% YoY, with passenger fare revenue up 849.3% YoY. This is a step toward turning "can driverless cars actually make money" from narrative into a line item on the balance sheet — Robotaxi now accounts for roughly 30% of the revenue mix and is growing far faster than the overall business. Operationally, the Robotaxi fleet expanded to 1,975 vehicles, with the company targeting over 3,500 by year-end; registered users on its China-based PonyPilot app surpassed 1.5 million. Internationally, the company says it has secured partners with an overseas Robotaxi deployment pipeline of over 4,000 vehicles, including an agreement with Uber to deploy 2,000 vehicles in Europe — this reflects planned/contracted figures, not yet delivered. The company also positions PonyWorld 2.0 as an R&D efficiency tool, saying it can scale deployment without a proportional increase in engineering resources. Source: Quiver Quantitative (via GlobeNewswire release) source
Tesla Cybercab could open to riders in Austin as soon as this month · autonomy ⚠️ unconfirmed report
According to The Information, Tesla has told employees it is preparing a public rollout of Cybercab in Austin, possibly as soon as this month. The plan calls for employees to first ride on public roads, with Cybercab formally entering Austin's Robotaxi service queue a few days later; groundwork over the preceding weeks has included test drives, private-road employee rides, and joint training with local emergency services. Cybercab has no steering wheel or pedals and is the vehicle Tesla's Robotaxi fleet plan actually depends on; production versions began road testing in June, with production ramp-up expected later this year. Tesla did not respond to a Reuters request for comment, and the timeline remains subject to internal notices only. Source: TradingView (via The Information) source
Apollo Go covers 28 cities globally, fully driverless mileage exceeds 240 million km · autonomy
Baidu's Q2 earnings disclosed that Apollo Go now covers 28 cities globally, with cumulative autonomous-driving mileage exceeding 350 million km, of which fully driverless mileage exceeds 240 million km. On international expansion, Dubai has launched fully driverless commercial operations, with rides bookable via the Uber app; Hong Kong granted the platform its first local fully driverless testing permit and began testing at Airport Island, making it the world's first platform to conduct fully driverless testing under right-hand-drive, left-side-traffic rules; London launched open-road testing in partnership with Uber and Lyft; the company also signed a strategic partnership with Kazakhstan's TPH, and has been conducting public-road testing with Post Bus in Switzerland since June. Source: Sina Finance source
Sany's first electric autonomous mining trucks exported to South America, extending Chinese mine-site autonomy into Latin America · industrial
Sany Group's first batch of SKT110Ei electric autonomous mining trucks shipped overseas from its Shenyang facility, marking the first autonomous mining-truck project to land in South America, with delivery covering the vehicles, dispatch system, and maintenance service. Backing this export is an already-proven operational base: as of July 2026, Sany's smart-mining unit has cumulatively deployed over 300 autonomous mining trucks, with commercial operations covering large open-pit mines in Inner Mongolia, Qinghai, and elsewhere in China, having surpassed 13 million cumulative safe operating km and moved over 41 million cubic meters of earth and rock. Mining is one of the few closed environments where autonomous driving has already achieved scaled, paid deployment, and selling the full system along with maintenance to an overseas mining company is a stronger test of technology maturity than a domestic demonstration project. Source: China News Service Hunan source
Kodiak launches autonomous truck testing on California public roads · autonomy
Kodiak AI will launch autonomous testing on California public roads, which the company says marks the starting point toward driverless coast-to-coast long-haul trucking. Under California's new rules, testing is unrestricted by location or mode, but a commercial deployment permit requires heavy trucks to complete a two-phase, 1 million-mile total testing program (500,000 miles per phase) and submit a structured safety case to regulators. The company is also rolling out its next-generation Gen7 platform to new fleets, with 50% more compute than the previous generation. On the customer side, Atlas Energy Solutions' Kodiak fleet will grow from 28 trucks to 100 by mid-2027; within a few weeks the company will also launch its first international pilot in Alberta, Canada, hauling logs for West Fraser Timber. Source: Trucking Dive source
Einride adds 500 Tesla Semi trucks, tripling fleet size · autonomy
Swedish electric and autonomous freight company Einride will purchase 500 Tesla Semi trucks, to be delivered in phases over 24 months starting in September, offered to customers including Amazon via its Saga AI fleet management platform. Einride currently operates roughly 200 owned heavy trucks, so the deal will triple its fleet size, covering key freight lanes in California, Georgia, New Jersey, and Texas; the company expects to convert roughly $800 million in potential long-term annual recurring revenue under joint commercial plans with shippers into actual revenue. The variable is on the supply side: Tesla only began high-volume Semi production in Nevada this past April, and in its Q2 shareholder letter walked back the term "mass production" for 2026, saying it first needs to ramp 4680 battery cell capacity. Source: TechCrunch source
4. Industry Developments
LG partners with Nvidia to amass 100,000 hours of humanoid robot training data by year-end · humanoid
LG Electronics is accelerating its partnership with Nvidia, targeting the generation of 100,000 hours of robot training data by the end of 2026 — by one converted estimate, equivalent to nearly 12 years of continuous operation. The data is produced at a robot data factory in Yangjae, Seoul, to refine LG's robot foundation model; LG deploys its CLoiD robots at the facility, collecting data across multiple simulated environments replicating homes, manufacturing, logistics, and manipulator settings — one zone is modeled on LG's washing-machine factory in Tennessee, US, where robots handle parts, stack them, and complete assembly. The collected data is first processed through Nvidia's data augmentation and synthetic-data generation before being fed back into training. This move comes four days after LG Group Chairman Koo Kwang-mo and Nvidia's Jensen Huang signed a physical AI cooperation memorandum at Nvidia's headquarters; Nvidia's senior director of Omniverse and robotics product marketing, Madison Huang, visited the Yangjae facility this past Tuesday to review progress. LG Electronics CEO Ryu Jae-cheol said the company will integrate group-wide capabilities under a "One LG" approach, aiming to become a comprehensive robotics solutions provider. Source: ANI/The Korea Herald source
Integral AI reaches the end of the road, exposing the funding gap in physical AI · world-model
Founded by two former Google researchers, Jad Tarifi and Nima Asgharbeygi, Integral AI built foundation world models for robotics and autonomous driving but raised only about $4.7 to $5.5 million in total seed funding, from investors including SoftBank's Deepcore and Samsung Next; the team numbered around 15 and had partnered with Denso since 2021 on industrial-robot skill acquisition, with Toyota and Sony also on its outreach list. In December 2025 it claimed to have launched "the world's first AGI-capable model," and as of March was still seeking about $10 million in a new round. Despite a strong founder pedigree, industry partnerships, and a differentiated technical approach, it still failed to secure follow-on funding — this is precisely where physical AI diverges from pure software AI: training data must be earned repeatedly through real-world interaction or fine-grained simulation, each data point costs more, failure consequences are heavier, burn rates are higher, and revenue paths are murkier — especially for companies building foundation models. Source: Crypto Briefing source
World Robot Conference 2026 opens tomorrow, with over 150 product debuts · industrial
The World Robot Conference 2026 will be held August 19–23 in Beijing's Yizhuang district, under the theme "Human-Machine Symbiosis, Coordinated Production and Demand." Over 300 companies will exhibit, with more than 2,000 products on display and over 150 product debuts; participants include Unitree, UBTECH, Galbot, Star Era, DEEP Robotics, and Fourier, among others. The main forum will release a "2026 Humanoid Robot Industry Development Report" and launch a "Global Robot Application Exploration Program," recruiting mass-production humanoid, quadruped, and dexterous-hand products to be made available free of charge to global innovation teams for real-world trials; a World Model Expert Committee under the China Institute of Electronics will also be established. Driven by this news, China's A-share robotics sector index rose 3.44% on August 18, with Estun Automation hitting the daily limit up. Source: CNFOL source
Second World Humanoid Robot Games begins competition August 22, with 2,056 robots participating · humanoid
Main competition runs August 22–26 at the National Speed Skating Oval, featuring 1,301 matches across 51 events, with 666 teams from 16 countries across six continents fielding 2,056 robots — team count up 138% and robot count roughly quadrupled from the first edition. The 30 athletic events add high-intensity contact disciplines such as long jump, weightlifting, tug-of-war, and table tennis, directly testing joint motors and reducers; the 21 scenario-based events extend into industrial, hospitality, home, and logistics settings, with a dedicated dexterous-hand category featuring 8 "micro-manipulation" subevents such as bean-picking and screw-tightening — upgraded from solo runs last year to head-to-head pairs this time. The dexterous-hand event kicked off early on August 18, with events such as nail-driving already underway. Source: Beijing Daily source
Chengdu unveils "Gongga No. 1" fully autonomous task-execution system · humanoid ⚠️ company claim
The Chengdu Humanoid Robot Innovation Center unveiled a fully autonomous (no remote control), long-horizon robot task-execution system operating under complex natural-language instructions, mounted on its ultra-lightweight humanoid robot "Gongga No. 1"; the related technology has been granted an invention patent. The center says the system integrates world models, VLA, and other technical approaches into a unified framework, allowing the robot to autonomously complete three sequential subtasks from a single complex instruction, with no remote control and no staging. "World's first" is the developer's own characterization, without independent third-party reproduction or standardized benchmark support. Source: Guandian.cn source
Pudu Robotics launches AI-native pallet-handling robot PUDU MP2000 · industrial ⚠️ company claim
Pudu Robotics launched the flat-ground handling forklift PUDU MP2000, designed to skip site modification and reflector-post deployment for mapping and go directly into production, while being directly compatible with standard pallet types such as Chinese "chuan" and "tian" character pallets. The company claims fork-pickup time is reduced by over 50% and cargo damage rate reduced by over 90% versus conventional approaches — both self-reported figures. On the product line, MP2000 complements the company's existing T-series AMRs, which focus on flexible delivery, extending Pudu's industrial-logistics coverage from standard load carriers to pallet handling. Notably, global penetration of unmanned forklifts remains below 2%, with deployment cycle and modification cost being the main barriers. Pudu has shipped over 130,000 units cumulatively, with business covering 85-plus countries and regions. Source: TOM News source
Unitree to list on STAR Market tomorrow; 9.78 million subscription accounts sets new sector record · humanoid
Unitree Robotics will list on Shanghai's STAR Market on August 19 (as previously reported). New today is IPO subscription data: valid online subscription accounts reached 9.7846 million, a new STAR Market record; the offering totals 40.4464 million shares, 10% of post-offering total share capital, but after strategic placement only 6.471 million shares remain for the online tranche — at 500 shares per lot, that's only about 12,900 lots — pushing the allocation rate down to 0.0181% given the supply-demand mismatch. Also notable is a talent-development move: on August 16, Unitree and the National AI Application Pilot Base (Embodied Intelligence) launched the country's first robot (embodied intelligence) vocational skills training center, under a "state-capital investment + Unitree technology enablement + professional operator" model, with Unitree providing training equipment, curricula, and instructors — targeting the shortage of frontline roles like operations, commissioning, and field adaptation, since even fully built robots can't scale deployment without people able to install and maintain them on-site. Source: Sina Finance source
Hardware · Supply Chain
· Dexterous hand pricing: unit prices have fallen from the RMB 50,000 range to under RMB 10,000, with the cost curve dropping fast source
· Sensortouch (triboelectric tactile sensing): technology originating from Southern University of Science and Technology's Ultra-Flexible Electronics Lab, recently secured bulk orders for dexterous-hand tactile modules; its fingertip 3D force module can decouple normal and shear forces, with minimum detectable force of 0.05N (about 5 grams-force) and response time of 10ms, using Infineon's PSoC 4 capacitive sensing chip for high-density synchronous acquisition source
· National Technology (300077.SZ): responding on an investor-interaction platform, said it has developed deployable solutions for joints, dexterous hands, and low-level motion control, with multiple products undergoing integration validation at several leading robot makers and some projects already deployed; the company also noted the industry has not yet reached large-scale volume and the corresponding business remains small — there is still a gap between upstream chip-side "integration validation" and actual shipment volume source
Top comments (0)