Today's Highlights
· Waymo sets end of 2027 for opening robotaxis to the public in Munich, its first entry into the EU
· China's Ministry of Industry and Information Technology lists intelligent robotics among six emerging pillar industries, with embodied intelligence named a future industry
· Yikong Zhijia (Chinese autonomous mining truck company) now operates over 3,100 driverless mining trucks, double a year ago
· Figure launches crowdsourced data-collection app, covering 108 countries with 16 million videos uploaded so far
· NVIDIA Jetson Orin Nano 2 doubles inference performance while cutting power draw 40% at equal performance
· Generalist's valuation rises to $3 billion, with 8VC leading a $200 million round expansion
· TrAct uses visual trajectories as an interface, lifting real-robot success rate from 49% to 76%
Research Progress
E-skin splits into fast and slow paths: familiar touch gets a millisecond reflex, uncertain touch gets handed to a language model for slow thinking · perception
The skin itself decides whether touch should be handled by reflex or by reasoning. Familiar contact is classified in milliseconds via spike-based encoding, while uncertain contact is routed into a language-guided reasoning pathway, with confidence determining which path is taken. An accompanying commentary by Wenzhuo Wu of Purdue notes that this confidence-based routing design balances speed and reliability in unpredictable environments.
Sun et al. · Nature Sensors (published online August 26, 2026) · Commentary source
TrAct treats visual trajectories as an intermediate interface for control and prediction · world-model
Robot actions are tied to the robot's own body and don't align with changes in image space — a longstanding problem for conditioning signals in world models. TrAct has the VLA model output candidate actions together with corresponding visual trajectories; a world model then simulates the visual outcome of each trajectory, and a vision-language reward model picks the one that best matches the instruction for execution. Against the π0.5 baseline, LIBERO-INTEGRAL success rate rose from 27% to 55%, and real-robot Franka manipulation rose from 49% to 76%. Li Fei-Fei is among the authors.
Zhi Cao et al. · arXiv 2608.24101 source
Do world models really follow the action? Swap in non-expert trajectories and the cracks show · world-model
Using a world model as a learned simulator assumes it produces a faithful future for any valid action — an assumption that has never been tested. WorldEcho probes this using visual integrity and SE(3) trajectory alignment, finding that existing models handle expert actions reasonably well but, when fed non-expert trajectories, either ignore the instruction or generate visually implausible outcomes. A companion method, WorldSync, addresses this via distribution coverage, representation grounding, and intervention-effect alignment, yielding higher success rates for policy iteration on RoboTwin and real robots.
Sixiang Chen et al. · arXiv 2608.24885 source
NVIDIA builds a surgical world model that humans can control in real time · world-model
Generative surgical simulation has previously been unable to run in real time. Cosmos-H-Dreams distills a bidirectional teacher into a causal few-step student via Self Forcing, achieving streaming inference at roughly 160 FPS on a single RTX PRO 6000 Blackwell; it is control-agnostic, drivable directly from a browser keyboard, a Meta Quest, or CMR Surgical's Versius surgical console, and learned policies can also run in closed loop. The team calls it the first interactive surgical world model supporting real-time control by both humans and policies, and it has been released as an open simulation system.
Javier Gamazo Tejero et al. (NVIDIA) · arXiv 2608.24199 source
Moving "imagination" off the critical path: a 48-millisecond action chunk that's more accurate than the synchronous version · world-model
Generating video synchronously can't keep up with control frequency, but dropping imagination altogether hurts success rate. GlanceWAM splits the two jobs inside a single video DiT: an asynchronous proposer runs on a slow clock in the background, glancing at a single future frame seconds ahead, while the action head decodes purely in latent space in 48-millisecond chunks without being blocked. Across 24 RoboCasa kitchen tasks it reaches 72.2%, above the synchronous Cosmos Policy's 67.1% and the no-imagination co-trained baseline's 64.4%, and reaches 99.0% on LIBERO at 24x the speed of the synchronous baseline. Code has been released.
Linhan Wang et al. · arXiv 2608.23927 source
Inference latency can break standard RL entirely — this asynchronous scheme fixes it · vla
Larger VLA models are slower, and slowness changes the environment's effective dynamics, which breaks the Markov assumption that RL relies on. ARLI augments an asynchronous inference framework with two additions: it folds the already-committed action and a mid-inference observation into the state, restoring an approximate Markov structure. Across simulated and real-robot manipulation tasks, standard RL fails completely under latency, while ARLI not only fine-tunes successfully but matches or exceeds standard RL under an idealized latency-free setting.
Brian Zhu et al. (Siemens · UC Berkeley) · arXiv 2608.23831 source
VLA models assume every gripper is the same — 103,000 demonstrations disprove it · vla
Parallel-jaw grippers and suction cups require different strategies to achieve the same grasp goal, yet existing VLA datasets are overwhelmingly parallel-jaw. MiGA spans multiple robots covering five gripper types and collects 103,000 demonstrations, explicitly recording strategy divergence for the same goal; the accompanying GVLA uses a multi-gripper tokenizer plus adapter-based policy routing, outperforming baselines on zero-shot and few-shot adaptation to new objects and unseen tasks.
Hanyi Zhang et al. · arXiv 2608.24603 source
Vision-tactile fingertip durability improved two orders of magnitude — swap the cartridge when it wears out · perception
Commercial GelSight Mini and DIGIT sensors ruptured after just 24–30 seconds in an abrasion test, and after 25–35 minutes under repeated compressive loading. This fingertip covers the silicone layer with an unpigmented, textured TPU protective film paired with a replaceable sensing cartridge, taking about 2–3 hours to reach rupture under the same conditions; under repeated loading at 39.2 newtons and 45 cycles per minute, nine test samples all remained functional when testing was stopped, with seven lasting the full 5 days. Damage is gradual and causes minimal interference with tactile imaging at end of life. Edward Adelson is among the authors.
F. Richard Cottrell et al. · arXiv 2608.24242 source
Other papers today: CARO embeds a fixed-base Euler-Lagrange model into the RL loop, achieving zero-shot robust quadruped control without torque sensors (arXiv 2608.24217 source); a fiber-optic sensing glove uses multi-core shape-sensing fiber to reconstruct full-hand pose, with fingertip error of 7.2mm dropping to 4.9mm after factory calibration (arXiv 2608.24572 source); Concept-Guided Exploration builds sustainable, executable scene graphs and selects hazard-mitigating actions in real time, 6 citations (arXiv 2608.23650 source); a survey of first-person intelligent smart-glasses platforms proposes an L0–L5 grading scheme, HF↑8 (arXiv 2608.24877 source); NeoWorld-Pro reframes monocular reconstruction as procedural generation of interactive 3D scenes (arXiv 2608.24212 source); CAT proposes trajectory-level continuous action representations, decoupling action representation from control frequency (arXiv 2608.24111 source); PonderPounce directly reuses an MLLM's native causal context as robot memory (arXiv 2608.24115 source); GaussianWAM distills the geometry and semantics of 3D Gaussian fields into world-action models (arXiv 2608.24714 source); Latent Action as Intention replaces explicit future-observation generation with latent actions, cutting WAM latency (arXiv 2608.24882 source); SIREN-Bench generates and evaluates emergency-vehicle interaction scenarios by behavior (arXiv 2608.24094 source).
Open Source · Tools · Benchmarks
· KUAVO-VLA: LejuRobot (Chinese humanoid maker) open-sources an industrial-vertical VLA that adds a layer between a general foundation model and task-specific skills. Fine-tuned on LingBot-VLA 2.0 using over 600 hours of same-embodiment real-robot data, it achieves 48.27% overall success across 25 tasks (20 custom industrial tasks + 5 GM100 tasks), versus 16.27% for the base model; π0.5, GR00T N1.7, and LingBot-VLA 2.0 were compared when selecting the base model source
· Pipette: Liquid AI open-sources an on-device model benchmark suite under Apache 2.0, treating "model × quantization × runtime × device" as a single unit of measurement. The initial dataset covers over 1,000 configurations and 30+ models, running llama.cpp builds on macOS/iOS/Windows/Android, validated on an M5 Max MacBook Pro, iPhone 17 Pro, and Galaxy S26 Ultra, with methodology independently reviewed by Artificial Analysis source
Funding & Deals
Generalist (USA) | Series B extension | Nearly $200 million | $3 billion post-money · world-model ⚠️ unconfirmed report
Following yesterday's disclosure of $200 million in new funding, the investor and valuation were confirmed today: 8VC led the round, expanding the $400 million Series B led by Radical Ventures in June at a $2 billion valuation to $600 million total, with regulatory filings confirming this is an extension. The company was founded in 2024 by former Google DeepMind researchers Pete Florence and Andy Zeng, and former Boston Dynamics engineer Andrew Barry, building robot foundation models that can drive different robot bodies, pairing off-the-shelf robotic arms rather than developing its own humanoid. The amount and valuation come from two people familiar with the matter; the company has not confirmed them publicly. In the same space, Physical Intelligence is valued at about $11 billion and Skild AI at $14 billion. Source: Tech Funding News source
GenRobot (Jianzhi Robotics) | Series A | Amount undisclosed | Over $200 million raised to date · adjacent ⚠️ company account
Momenta led the round, with existing investors following on. This company, which builds embodiment-agnostic data and evaluation infrastructure, is only a year old and has already completed 7 rounds. It has developed the Gen DAS unobtrusive wearable data-collection device and the Gen Controller skill terminal, recording human operations via sound, magnetic tactile sensing, piezoresistive sensing, capacitive sensing, flexible fabric, and force feedback across multiple channels; its self-trained Data Foundation Model covers pretraining dimensions including chain-of-thought, 6D pose, depth, tactile sensing, hand tracking, and full-body mesh. The company says it already has embodied models pretrained entirely on its embodiment-agnostic data, demonstrating cross-scenario transfer in complex manipulation generalization tests. Source: CYzone source
Futuring Robot | New round | Over RMB 1 billion raised across three rounds in half a year · embodied ⚠️ company account
This round drew nearly 10 investors, including ByteDance, CATL Capital, Guofang Venture Capital, Deneng Capital, and Nice Group, with existing investors also increasing their stakes. Founded in 2022, the company builds general-purpose household robots for consumers and has launched two models, F1 and F2; in May it released its new-generation robot brain, Self-Evolving WAM, and deployed it to F2, which at WRC 2026 completed a seven-step meat-sauce pasta task end-to-end. The company says its robots are now in regular operation in over 500 ordinary households nationwide, with a core technical team of nearly 200 people. Source: Gasgoo source
MileBot | Strategic financing | Nearly RMB 100 million · adjacent
Funded by CITIC Securities, with Haoguan Capital acting as exclusive financial advisor. Founder Chen Gong holds a PhD in biomedical engineering from the National University of Singapore, and the company's approach uses self-developed compliant actuators for active rehabilitation training, rather than rigid exoskeletons that passively move the patient. 36Kr calls it the first company in China's medical exoskeleton sector to achieve a working profit model. The company cites clinical validation from the First Affiliated Hospital of Nanjing Medical University and Shenzhen Second People's Hospital: stroke patients in the active-training group showed significant improvement over baseline by week 2, versus week 4 for conventional treatment. Next steps focus on brain-computer interfaces and home elder-care scenarios. Source: 36Kr source
SYNAPATH AI | Angel round | Nearly RMB 100 million · adjacent ⚠️ company account
Builds human behavior data products spanning Ego, Exo, Tactile, and other data modalities, with a pipeline running from collection, cleaning, and motion/spatial information extraction through to quality inspection and delivery. The company says it has accumulated several hundred thousand hours and over 36 million data entries, spanning more than 260 scenario types, and has partnered with embodied-AI companies including AgiBot, Fourier Intelligence, and Robot Era, as well as major internet firms, on data and skill transfer. Source: Gasgoo source
Lingyu Intelligence | Pre-A round | Several hundred million RMB · adjacent
Jointly invested by Zhuzhou Industrial Investment, Feitu Venture Capital, Future Boundary Venture Capital, and Xinneng Venture Capital — the company's third round this year. Founded in early 2025, it develops the Lingyu TA robot platform in-house, paired with a real-robot data-collection pipeline and cloud-based operations platform; the funds will go toward expanding mass-production capacity for the platform and upgrading the data-collection pipeline. The company says it has already achieved large-scale commercial deployment in open retail settings. Source: PEdaily source
Commercialization & Deployment
Waymo sets end of 2027 for opening robotaxis to the public in Munich, its first entry into the EU · autonomy
In the coming weeks, Waymo will begin phased road testing in Munich, with human drivers mapping routes, training autonomous specialists, and adapting the software to local driving habits, ahead of a public launch planned for the end of next year. This marks Waymo's third international city after London and Tokyo, and its first entry into the EU. It will start with a small all-electric fleet of Jaguar I-Pace vehicles before expanding. Co-CEO Tekedra Mawakana said Munich is a world-class hub for mobility and engineering, that earning community trust ahead of launch is the top priority, and that the company will invest in local fleet operations and create high-skill jobs. Waymo currently operates in 11 U.S. cities, with 19 more in progress; as of March this year, it had logged nearly 221 million fully driverless miles. Last month, Uber also announced a partnership with Pony.ai and Croatia's Verne to launch service in Europe, with Munich also on its planned list. Source: CNBC source, Silicon Republic source
Yikong Zhijia's driverless mining truck fleet tops 3,100, double a year ago · autonomy
The strongest number in the interim results wasn't revenue. As of June 30, the company had over 3,100 driverless mining trucks in operation, up from over 1,600 a year earlier, with more than 1,000 additional trucks ordered and in delivery. By the company's own research, China's total operating fleet is about 6,000 trucks, giving it a share of over 53%. It serves 38 mines, up 12 from a year earlier, including 9 mine sites with fleets over 100 trucks and 4 over 200, with the largest single mine at 566 trucks. Overseas, in a partnership with Zijin Mining's Norton Gold Fields and Thiess in Australia, the first 6 mining trucks entered real production at the Kalgoorlie Havana open-pit gold mine in Western Australia. Financially, the picture diverges: revenue was RMB 548.8 million, down 17.3% year-on-year, while gross margin rose from 1.7% to 18.9%, and net loss narrowed 20.2% year-on-year to RMB 188.8 million; the share of revenue from the customer-provided-fleet model rose from 62.7% to 65.6%. Source: Landsea Green Life source
Camera towers take over receiving inspection at a cold-storage facility, logging 64,000 receipts in six months · industrial
At Lineage's cold-storage facility in Decatur, Alabama, receiving has been handed to Kargo's fixed camera towers. As pallets enter the dock, they are imaged from multiple angles, and computer vision extracts label text, barcodes, batch numbers, expiration dates, and dimensions, writing directly into the warehouse management system, with damage or mismatched paperwork flagged in real time with photo evidence. The facility processes over 500 pallets a day; a task that used to take several minutes per receipt now takes seconds, and over 64,000 receipts have been logged in the past six months. Lineage's VP and Chief Data Scientist Elliott Wolf said automation frees up staff for higher-value work rather than replacing them. The hardware is rated to withstand forklift impacts, plugs into standard 110–120V outlets, and doesn't require Wi-Fi. Automated outbound loading scans are also set to launch at Decatur. Source: Unite.AI source
Chengdu-made L4 delivery robots hit European streets · autonomy ⚠️ planned figures
Yangsi Technology (Chengdu-based robotics company), whose mass-produced L4 autonomous logistics robots number in the thousands, has these robots already running regularly on public roads in several overseas cities, handling food delivery, retail, and last-mile transfer orders. The units have over 60 liters of cargo capacity, a 20kg payload, support fast battery swaps, offer over 12 hours of range per charge, use 6-wheel drive with independent suspension, can clear 14cm obstacles, reach a top speed of 10km/h, and sense their environment via lidar, cameras, and ultrasonic radar. The company plans to deploy 20,000 units across an entire overseas country by the end of 2027. Source: Sohu source
XPeng's robotaxi completes over 2,000 internal test rides in Guangzhou · autonomy ⚠️ earnings-call account
He Xiaopeng said on the Q2 earnings call on August 24 that the production robotaxi equipped with second-generation VLA has completed over 2,000 internal test rides in Guangzhou, validating the full passenger-demonstration workflow, with the cloud-based remote takeover platform now complete; the target is to begin carrying passengers without a safety driver in 2027. The second-generation VLA has already completed localized testing in Germany, aiming for one model to handle both Chinese and European road conditions. Q2 total revenue was RMB 19.74 billion, up 8.0% year-on-year, with 103,295 vehicles delivered and R&D spending of RMB 2.91 billion, up 32.1% year-on-year. Source: Gasgoo source
Industry Developments
Intelligent robotics enters the "15th Five-Year Plan" as one of six emerging pillar industries; embodied intelligence named a future industry · adjacent
At a State Council Information Office briefing, Vice Minister of Industry and Information Technology Xin Guobin laid out the industrial structure for the next five years: accelerating the buildout of six emerging pillar industries — integrated circuits, aerospace, biomedicine, the low-altitude economy, new-type energy storage, and intelligent robotics — while advancing quantum technology, biomanufacturing, hydrogen energy and nuclear fusion, brain-computer interfaces, embodied intelligence, and 6G as future industries and new growth drivers. The AI section covered four areas: supply, application, ecosystem, and safety. On the supply side, it named support for R&D of high-end training chips and multimodal algorithms, tackling frontier technologies such as brain-inspired intelligence and world models, and advancing intelligent terminals such as humanoid robots and brain-computer interfaces. On the application side, it proposed a "one industry, one map; one scenario, one profile" approach to scenario cultivation, aiming to distill decades of manufacturing know-how into lightweight, low-cost, easily deployable industry solutions. Yao Jun, director of the Planning Department, gave quantified targets: by 2035, the goal is to cultivate roughly 100 demonstration parks and 1,000 demonstration enterprises. The new-energy-vehicle section noted plans to actively and steadily advance pilot programs for intelligent connected vehicle access and road use, and for vehicle-road-cloud integration. Source: Guangzhou Daily / Xinhuacheng source
First dedicated dexterous-hand competition shifts the evaluation bar from "can it run and jump" to "can it do the work" · embodied
Eight competition events covered fine manipulation tasks such as powder weighing, tweezer bean-picking, cable connection, and tool assembly, with an average of 51 teams registering per event and only 7 to 17 advancing to finals. AgiBot's subsidiary Lingjiedian (Critical Point) entered its mass-produced dexterous hand, OmniHand, for the first time and swept awards in all 8 events, taking 7 gold, 4 silver, and 3 bronze; all 14 medals came from the same unmodified production unit — a 510-gram, 16-degree-of-freedom hand. The competition scored teleoperation and full autonomy separately, weighting autonomous performance at 1.0 versus 0.5 for teleoperation, though most teams still chose teleoperation. Among Critical Point's 3 autonomous gold medals, the powder-weighing win used algorithms developed independently by a Wuhan University team on OmniHand hardware. Another team, Lingxin Qiaoshou (Dexterous Hand), was the only team in the competition to compete fully autonomously across all eight events, driving in 18 screws in 5 minutes during the power-tool assembly final, missing points on only one screw for slightly exceeding torque tolerance, scoring 17 out of 18 to lead second place by 8 points, using an 11-degree-of-freedom, 370-gram Linker Hand O6. In the scenario events, Galbot (Chinese robotics company) competed fully autonomously using its self-developed AstraBrain world-action model, sweeping gold in the household, dining, and retail categories, finishing the household event with 305 points and the fastest completion time. Critical Point CEO Qiao Tianjie drew a line in an interview: for tasks involving impact loads and multi-step planning, such as fixed nail-driving or unpacking, they still had to fall back to teleoperation. Source: Sohu source, source, PR Newswire source
Figure crowdsources data collection worldwide, app now live in 108 countries · adjacent ⚠️ company account
On August 25 U.S. time, Figure launched Index simultaneously on iOS and Android, paying users worldwide to submit real first-person videos of manipulation tasks. It now covers 108 countries and regions, with over 264,000 downloads, more than 44,000 weekly active contributors, and over 16 million videos uploaded. The company says it will invest over $1 billion in data and compute over the next 12 months. Figure had previously relied on third-party data vendors, but found that scale, diversity, and quality couldn't keep up with what training its Helix model required.Source: 36Kr source
Jetson Orin Nano 2 doubles inference performance while cutting power draw 40% at equal performance · hardware
The module delivers 78 TOPS of compute, 8GB of memory, and an 8-core Arm CPU, keeping the same module form factor as the previous-generation Orin Nano Super. NVIDIA says the gains come from an improved Tensor Core and higher memory bandwidth; in 15-watt mode it matches the Super's performance while using 40% less power. On the software side it runs Jetson AI Lab's agent skills, supporting Cosmos, Nemotron, and edge-tuned models such as Gemma 4 and Qwen 3, with NVIDIA saying over 3 million developers have used its robotics stack. Deepu Talla, VP of Robotics and Edge AI, said, "today's small and medium frontier models have reached the accuracy of last year's largest frontier models, unlocking real-time intelligence for edge devices." Among disclosed adopters, Alphabet's drone delivery unit Wing is evaluating the module, while home-cleaning robot company Matic has already adopted it to run conversation, gesture detection, and semantic fine-grained mapping locally; Cognex and Doosan Bobcat are also on the list. Aptiv separately announced the same day that its production-grade platform supports the module. Source: Unite.AI source
XPeng G9L debuts second-generation VLA, with X-Foresight predicting 6-second trajectories · autonomy
An XPeng vice president announced today that the G9L will be the first vehicle to feature a new version equipped with second-generation VLA, with its accompanying X-Foresight world-prediction model able to project environmental motion trajectories 6 seconds into the future. He Xiaopeng said earlier on the earnings call that the second-generation VLA model will get its first major upgrade at the end of August, with the on-device model's parameter count increasing 3.5x. The vehicle opened for pre-order on August 11 at a pre-sale price of RMB 259,800, available in super range-extender and pure-electric variants, with 800V high-voltage ultra-fast charging across the lineup. A Physical AI sharing and experience event is scheduled for August 27 at 15:00. Source: Sina Technology source
5-year-old girl loses 4 teeth after being kicked by a performance robot; operator confirms it was a Unitree product · humanoid
At a science museum in Xiangyang, Hubei, a 5-year-old girl watching a robot performance was struck in the face by the robot's raised leg, losing 4 teeth and requiring 3 stitches to her face. The incident occurred in November 2025; the museum said the operator "failed to anticipate the safe distance and misoperated the robot." China Newsweek confirmed with the operating company, Andar Innovation Exhibition, whose marketing manager stated explicitly that the robot was a Unitree product. Similar incidents have occurred elsewhere: in June 2026, a humanoid robot performing martial arts in Urumqi struck a boy in the abdomen with a spinning kick, and in February an approximately 1.6-meter-tall robot at Chengdu's CapitaMall deviated from its preset path and knocked over an elderly person; neither site had physical barriers separating the robots from the public. The market itself is expanding — iiMedia Research data shows China's robot rental market was worth about RMB 1 billion in 2025 and is projected to exceed RMB 10 billion in 2026, with Orient Securities giving a similar-magnitude forecast for the performance-robot rental market specifically. One amusement-park operator told the magazine, "the barrier to entry is genuinely low — if you can afford one, you can run this business"; the venue keeps 4 dedicated operators on staff and requires crowds to be cleared before major movements, but acknowledged the industry has no unified standards — "all of this relies on self-discipline." The family has repeatedly requested the operator's qualifications and training records, none of which have been provided to date. Source: iFeng source
Qiongchu Intelligent releases Noe-0, an entirely teleoperation-data-free pipeline · world-model ⚠️ company account
Using no teleoperation data at all in either pretraining or post-training is this world-action model's most aggressive design choice. The team used its self-developed RoboPocket collection device, letting operators perform tasks directly by hand in real homes and stores, gathering self-collected data spanning over 50 cities, several hundred thousand hours, and hundreds of thousands of task types. One of the reasons the team gives is practical: completing the same task via teleoperation typically takes 3 to 5 times longer than direct manual operation, and tends to produce slow, mechanical motions. Architecturally, the policy backbone is swapped from a language model to a video world model, using video prediction as the core high-level learning objective; the team says that even when there is a mismatch between the collection setup and the deployment robot body, pixel prediction still meaningfully improves cross-embodiment transfer. Technical details were published on the team's research preview blog; weights have not been open-sourced. Source: Sina Technology source
Kodiak puts AMD EPYC chips in its seventh-generation driverless truck · autonomy
In the seventh-generation Kodiak Driver, an EPYC CPU handles fusion and preprocessing of lidar, camera, and radar data while also supporting path planning. AMD's senior director of automotive markets, Wayne Lyons, said the key CPU workloads in autonomous driving are latency-sensitive and hard to parallelize. Kodiak founder and CEO Don Burnette said using off-the-shelf commercial hardware lets the company focus resources on building trucks that can be mass-produced. The company's driverless fleet in the Permian Basin now operates with no one in the cab. Source: The Manila Times source
Hardware · Supply Chain
· Standardized robot battery: JD.com launched a standardized battery series at WRC 2026, with cells supplied by BAK Battery, using unified interfaces and dimensions for plug-and-play compatibility across humanoid, industrial, and wheeled robot forms; BAK is also pushing a "one cell, three variants" strategy for humanoids, pairing a mature 21700 cell with a unified physical/electrical architecture, then offering burst-output (running, jumping, load-bearing climbing), endurance (commercial service), and balanced (home and industrial collaboration) variants source
· Silicon carbide for humanoids: UBTECH (Chinese humanoid robot maker) signed a strategic partnership with BASiC Semiconductor around SiC power devices, for joint research and validation on power-management and motion-control efficiency and endurance improvements in robots; BASiC Semiconductor plans to deploy UBTECH humanoid robots on its own silicon-carbide production lines for wafer fabrication and module packaging source
Top comments (0)