DEV Community

Shawn
Shawn

Posted on

FutureX · Physical AI Daily — Issue 51 (07/08)

Today's Highlights

· Official data: Jan–May above-scale robotics enterprise revenue tops RMB 90 billion; H1 embodied AI financing in China reaches RMB 93.5 billion, up fivefold year-on-year

· Hikvision Robotics' cumulative mobile robot production surpasses 200,000 units, doubling again in two years — global market share No. 1 in 2025

· Jijia Vision (Chinese world-model startup) raises over RMB 3.5 billion across three rounds in 3 months, as the world-model funding race continues

· Nvidia partners with Hugging Face to release GR00T N1.7 and other new models for open-source LeRobot

· AgiBot (Chinese humanoid robotics maker) brings its Yuanzheng A3 to Europe, launching Robotics-as-a-Service (RaaS) in the UK

I. Research Papers

GigaWorld-1: A roadmap for using world models to "score" robot policies · world-model

Robot policy evaluation has long been bottlenecked by the need for real-robot rollouts — slow and expensive. This paper systematically studies world models as scalable "policy evaluators": it first clarifies what properties a world model needs to credibly substitute for real-robot rollouts, then builds an evaluation benchmark called WMBench using real teleoperated robot data paired with policy replays, enabling controlled comparisons across model families and action-encoding schemes. It was the day's highest-trending paper (HF 29↑), making the case that "world model as evaluation infrastructure" is a real research direction.

GigaWorld Team et al. · arXiv 2607.02642 source

InternVLA-A1.5: Unifying understanding, latent foresight, and action into a single policy · vla

Unified manipulation models want to simultaneously capture the semantic priors of pretrained VLMs and the physical dynamics learned from future prediction. This version merges both with action decoding into a single framework, emphasizing compositional generalization — recombining skills, objects, and scenes seen in training into unseen tasks. Part of the InternVLA series from Shanghai AI Lab, HF 16↑.

Haoxiang Ma et al. (Shanghai AI Lab · InternVLA) · arXiv 2607.04988 source

The first "multiplayer" interactive world model: it can tell who caused what · world-model

Previous world models treated other agents as part of the environment. This is the first multiplayer world model built for complex physical interactions: conditioned on each player's individual action stream, it learns to attribute scene changes to the correct player — even when everyone accelerates, jumps, and flips simultaneously, the scene doesn't flicker or distort. It corresponds to the previously released MIRA demo (multiplayer, roughly 20 FPS real-time), HF 10↑.

Anthony Hu et al. · arXiv 2607.05352 source · Analysis: Zhineng Zhixin source (WeChat, CN)

Vision pretraining for dense spatial perception · perception

Modern vision foundation models favor "semantic invariance," often at the cost of spatial detail — but embodied AI needs structured, scale-aware, actionable spatial representations recoverable from pixels. This paper builds a pretraining method specifically for dense spatial perception, addressing a gap in mainstream vision models. Tied for the day's highest community trending score (HF 27↑).

Zelin Fu et al. · arXiv 2607.05247 source

TACO: Using a tactile world model as a VLA's "self-correction" mechanism · manipulation

VLAs often fail unrecoverably on contact-rich tasks due to small contact disturbances invisible to vision. TACO uses a tactile world model to turn real-world failures into "imagined visuo-tactile corrections" for post-training, and uses knowledge-isolated tactile adaptation to avoid corrupting pretrained vision-language priors.

Shengbang Liu et al. · arXiv 2607.02840 source

DREAMSTEER: Steering VLAs at deployment time, with no fine-tuning · vla

Pretrained VLAs generalize reasonably well zero-shot, but stumble when distribution shifts at deployment; the standard fix is fine-tuning on target-environment data — which requires having demonstrations in the first place. DREAMSTEER changes zero parameters, instead using a latent world model plus a value model to steer the policy toward more reliable actions in real time at deployment. Includes authors from Meta FAIR.

Hanchen Cui, Sergio Arnaud et al. (incl. Meta FAIR) · arXiv 2607.02865 source

CLEAR: Bringing large-scale closed-loop RL to end-to-end autonomous driving · autonomy

Most VLAs in end-to-end autonomous driving are open-loop imitation, which struggles to self-correct on long-tail scenarios. CLEAR performs large-scale closed-loop reinforcement learning, learning a residual waypoint policy on top of a pretrained VLA's waypoint priors — preserving the knowledge in the large model while letting trajectories be refined by real closed-loop feedback. From Qualcomm AI Research.

Yunxiao Shi et al. (Qualcomm AI Research) · arXiv 2607.02841 source

iFLYTEK-Embodied-Omni: iFLYTEK's unified embodied foundation model · vla

Aimed at general-purpose embodied agents, this jointly models vision, language, and action within a single Omni framework, emphasizing "brain–cerebellum" coordination: understanding multimodal instructions and anticipating environmental changes, while producing precise control over long horizons. A technical report from iFLYTEK, and another unified-foundation-model attempt from a major lab that day.

Yuan Zhang et al. (iFLYTEK) · arXiv 2607.02542 source

ADP: Adding "adversarial dynamics priors" for disturbance-resistant humanoid locomotion · locomotion

Existing motion-prior methods only imitate kinematic style, without directly constraining dynamic quantities like center-of-mass motion, momentum, and contact forces. ADP uses an adversarial dynamics prior to directly regularize these quantities, yielding humanoid locomotion control with stronger disturbance resistance.

Seokju Lee et al. · arXiv 2607.03454 source

Other papers today: ResTacVLA, using residual tactile representations to address "modality collapse" where touch gets drowned out by vision in VLA (arXiv 2607.03387 source); HiMe, hierarchical embodied memory tackling the "frequency-capability paradox" in long-horizon VLA (arXiv 2607.03449 source); Current as Touch, learning contact compliance from motor current and joint state alone, without external tactile sensors (arXiv 2607.03529 source); Cross-Embodiment, a unified hand-action space for cross-morphology transfer (arXiv 2607.03570 source); WorldBagel, using a unified multimodal model for vision-language-action-world modeling (arXiv 2607.03461 source); LLM-as-a-Verifier, treating "verification" as a new scaling axis (arXiv 2607.05391 source).

Open Source · Tools · Benchmarks

· Nvidia × Hugging Face · LeRobot: Introduces GR00T N1.7 and other new models plus a Teleop framework for the open-source robotics community LeRobot, lowering the barrier to open-source embodied development source

· EVA-Client: An open-source unified framework for real-robot policy deployment, data collection, and evaluation, bundling mainstream inference strategies like ACT temporal ensembling and Real-Time Chunking into one codebase (arXiv 2607.02646 source, HF 21↑)

· Deform360: A large-scale multi-view visuo-tactile dataset for world models of deformable objects (arXiv 2607.05390 source)

· RoboDojo: A general manipulation-policy benchmark unifying simulation and real-robot evaluation on the same scale (arXiv 2607.04434 source)

II. Funding & Deals

Jijia Vision | Pre-B/B1/B2 | Over RMB 3.5 billion combined · world-model

Three rounds in three months, bringing together state-backed and auto-supply-chain capital including Sino-CEE Fund, CCB Investment, Wanxiang Qianchao, and Fosun RZ Capital. Beyond its own hardware, the company is investing heavily in data-collection hardware and its world-model platform, pursuing the "simulate physics from video/data" generative approach. Following Shengshu and Kuawei, the world-model funding race keeps heating up — in the same period, Qianxun Intelligent raised over RMB 4.5 billion across 3 months, and Xingyuan Zhi (incubated by BAAI) has raised a cumulative RMB 1 billion.Source: Qingcheng Finance source (WeChat, CN)

LinkerBot | Planned Hong Kong IPO | Market-estimated valuation over RMB 100 billion · hardware ⚠️ Planned, unconfirmed

Focused on high-DOF dexterous hands for humanoid robots (Linker Hand series, 6–42 degrees of freedom); if it proceeds, this would be another dexterous-hand player sprinting toward public markets. The valuation is a market estimate, not yet finalized.Source: Yijian Finance source (WeChat, CN)

Tesollo (South Korea) | Launches IPO process after Series B · hardware

A South Korean dexterous-hand developer has begun preparing for an IPO after completing its Series B, focused on humanoid robot end-effectors. Alongside LinkerBot's IPO move, dexterous hands — the "highest-value component" in humanoids — are collectively heading toward public capital markets.Source: The Robot Report source

Kraken Robotics | Acquires Covelya Group | USD 615 million · autonomy

Marine robotics maker Kraken is acquiring Covelya for USD 615 million, the largest deal of the day, as consolidation in underwater autonomy accelerates.Source: The Robot Report source

Shihang Intelligent (Kunshan) | New funding round | Claims "largest-ever financing in the global marine robotics sector" · embodied ⚠️ Company's own claim

Nurtured from its angel round by the Kunshan Industrial Technology Research Institute, this round is open only to a small group of industry-resource investors and existing shareholders, with indicated subscription interest several times the target raise. "Largest globally" is the company's own characterization; the specific amount has not been clearly disclosed.Source: Suzhou Daily source (WeChat, CN)

Quick roundup: Small seed rounds in industrial/factory physical AI · industrial

Mowito (India) raised a USD 3 million pre-seed (led by Version One, All In Capital), teaching factory robots "by demonstration rather than programming" source; HIVE raised USD 15 million to build physical AI for industrial manufacturing source; Daxiao Robotics (part of SenseTime's "1+X" ecosystem) raised a cumulative several hundred million USD across two rounds in 4 months, and its Kaiwu world model Kairos claims to top 4 embodied benchmarks (⚠️ company's own claim) source (WeChat, CN).

III. Commercialization & Deployment

Hikvision Robotics rolls off its 200,000th mobile robot · industrial

Cumulative production across all mobile robot categories has surpassed 200,000 units — the second 100,000 took just two years, since hitting the 100,000th unit in May 2024. According to the Mobile Robot Industry Alliance, Hikvision Robotics held the No. 1 global market share in 2025. This is one of the day's few genuine mass-production milestones — compared to vendor demos and letters of intent, "units shipped" is a hard scale metric.Source: Eastmoney/Southern Finance source

Zerith Robotics lands its first overseas full-robot order, in Malaysia · humanoid

On July 5, three wheeled humanoid robots — the ZERITH-H1, made in Hefei — performed teleoperated grasping demonstrations at a lab in Malaysia, marking the company's first overseas order for complete units. The order is small and still early-stage deployment, but it extends Chinese humanoid delivery from domestic use to overseas settings.Source: Hefei Development and Reform Commission source (WeChat, CN)

Pony.ai's Singapore robotaxi service integrates with the Zig ride-hailing app · autonomy

Pony.ai's robotaxi service in Singapore is now bookable through the Zig app, connecting its fleet to a mature local ride-hailing entry point and moving toward regular, on-demand availability. Following WeRide's partnership with Grab last week, Singapore has become another hotspot for Chinese robotaxi players' overseas deployments.Source: TechRepublic source

Ambi Robotics integrates with Pickle Robot to automate warehouse inbound logistics · industrial

The two companies have connected their respective AI robots to automate warehouse inbound logistics (unloading and upstream sorting) — a segment of warehouse automation that has long relied on manual labor and is seen as the next scalable growth area.Source: Robotics & Automation News source

Simplexity Robotics: 11 months old, has delivered 100 robots to production lines · embodied ⚠️ Single source

This embodied-AI startup claims to have delivered 100 robots to production lines, touting the "fastest to 100 units" delivered. The actual scale is modest and the claim currently comes from a single media outlet — noted here pending independent confirmation.Source: Pandaily source

IV. Industry Developments

Fei-Fei Li publishes a "functional taxonomy of world models," drawing lines around an overused concept · world-model

Fei-Fei Li writes bluntly that "world model is now one of the most important — and most overused — terms in AI today": computer vision, robotics, reinforcement learning, and generative AI all claim to be building world models, yet mean very different things by it. She divides world models into three categories — rendering, simulation, and planning — which in practice further converge into the "simulate physics from video/data" generative approach (Jijia Vision, Momenta, etc.) and the "verify the world through action" end-to-end interactive approach (Zibianliang, Qianxun, etc.). This definitional debate comes against the backdrop of a "world-model boom" in China that has raised tens of billions of RMB in three months, and it's also why the day saw a cluster of analyses and cost-reduction announcements (Moxin claims a Flash model cutting costs 70% while running at 50 FPS; Amap's general-purpose world model reportedly reaches 16 FPS on 8 RTX 5090 GPUs).Source: Yanzhi Auto source (WeChat, CN)

AgiBot's Yuanzheng A3 lands in Europe; UK launch is first Robotics-as-a-Service (RaaS) offering · humanoid

AgiBot unveiled its Yuanzheng A3 humanoid robot in Europe and, together with ten local partners, launched a "Robotics-as-a-Service (RaaS)" rental model in the UK — upgrading its overseas push from selling hardware to selling a service. In China, its mass-production line has reportedly shipped its 15,000th unit (previously reported), and it signed a comprehensive strategic partnership with appliance maker ASD (Aishida), locking in actuator/lead-screw supply and application deployment. AgiBot's supply-chain executive said "China's supply-chain manufacturing capability will never become a bottleneck," predicting the embodied-AI inflection point will arrive within 2–3 years.Source: Weilai Keji Shihuangzhe source (WeChat, CN)

Momenta's autonomous driving system surpasses one million vehicles equipped · autonomy

Momenta says its ADAS/autonomous driving system has now been factory-installed in over one million vehicles, with partner automakers including Mercedes-Benz, BMW, and Audi. Technically, it bets on the "simulate physics from video" world-model approach rather than pure VLA — its first mass-production vehicle featuring the R7 world model — and founder Cao Xudong says "sensors rank last in importance." The million-unit figure is a genuine factory-installation number, adding another scale data point to the company's Hong Kong IPO narrative.Source: Ifeng Auto source

Ant Group's embodied-AI team releases LingBot spatial perception model · perception ⚠️ Company's own claim

Ant Group's robotics team has released LingBot-Vision / LingBot-Depth 2.0, claiming "12 world firsts" and that its 1.1B-parameter model outperforms the 7B DINOv3, targeting hard-to-perceive objects like transparent glass — reportedly significantly reducing robot collisions with glass. Both the leaderboard rankings and the "world first" claims are self-reported by the company, with no independent benchmark verification observed.Source: Pandaily source

Nvidia unveils full autonomous-driving lineup at GTC 2026, deepens open-source robotics ecosystem · autonomy

Nvidia presented a complete autonomous driving solution stack at GTC 2026, and on the same day partnered with Hugging Face to bring GR00T N1.7 and other new models and frameworks to the open-source community LeRobot (see "Open Source" section). The company is advancing on two fronts simultaneously — vehicle-side autonomous driving and an open-source embodied-AI foundation — as part of a dual-pronged platform strategy.Source: Autohome source

Huawei-trained talent spills into embodied AI en masse — 9 "Genius Youth" alumni companies have raised roughly RMB 7.7 billion combined · adjacent

According to a tally as of July 6, at least 15 former Huawei executives have founded or joined embodied-AI companies, including 9 alumni of Huawei's "Genius Youth" program. These companies have publicly disclosed a combined roughly RMB 7.7 billion in funding, producing two RMB-10-billion-plus unicorns — AgiBot and Galbot. Core Huawei teams like Ascend, the Automotive BU, and Noah's Ark Lab are becoming a "West Point" for embodied-AI talent.Source: Zhidongxi source (WeChat, CN)

Policy push intensifies: 6 national AI standards initiated, Guangzhou rolls out new embodied-AI incentives · adjacent

China has initiated 6 national standards covering large models, embodied intelligence, and AI safety, setting rules for a rapidly expanding industry source; Guangzhou has issued 8 measures to promote its embodied-AI robotics industry, with subsidies up to RMB 1 million per application scenario source (WeChat, CN). Together with dedicated real-world training initiatives from the Ministry of Industry and Information Technology and SASAC, an institutional "standards + subsidies" framework is taking shape.

Hardware · Supply Chain

· Dexterous-hand-dedicated chips: China's first integrated chip solution for dexterous hands has launched, claiming 0.1mm positioning accuracy and microsecond-level real-time latency, aimed at the compute bottleneck limiting dexterous hands' scale commercialization source (WeChat, CN)

· Magnetic encoders: Humanoid robots are driving demand for magnetic encoders; MPS (MiniIC, Chinese semiconductor firm) is moving into this emerging "sensory organ in the joints" niche of position sensing source

· Planetary roller screws: Schaeffler showcased humanoid-specific planetary roller screws, high-rigidity bearings, and planetary-gear actuators, redefining joint-drive standards through high-precision torque transmission source (WeChat, CN)

· Zhaowei Machinery & Electronics: Unveiled three new dexterous-hand products at electronica Munich in Shanghai; GGII data puts China's 2025 dexterous-hand sales at roughly 19,200 units, projected to exceed 430,000 by 2030 (roughly 87% 5-year CAGR) source (WeChat, CN)

· Chinese reducer makers tie into overseas leaders: Leaderdrive is currently the only Chinese harmonic-drive maker to enter Tesla Optimus's supply chain, with orders booked through 2027; Shuanghuan Driveline plans capacity of up to 500,000 units by 2026. Official figures claim domestic-content localization of core humanoid components has already surpassed 75% source (WeChat, CN)

Top comments (0)