FutureX · Physical AI Daily — Issue 103 (08/29)

Today's Highlights

· Anthropic releases hardware MCP, letting AI agents directly control lab instruments and production line equipment

· UBTECH (Chinese humanoid maker) sold 16,123 units of humanoids in H1, full-size model revenue up 1,445%

· a16z raises $1.1 billion "American Dynamism"-style fund focused on AI's physical infrastructure

· Unitree's G1 EDU exposed with two root vulnerabilities, one starting from Bluetooth

· China's National Development and Reform Commission cools robotics enthusiasm: prevent blind bandwagon-jumping and herd behavior

· Riemann-1.0 hits 85% long-horizon real-robot success rate, 15 points above the strongest open-source baseline

Paper Progress

Riemann-1.0: compressing robot policy and world simulator into a single causal autoregressive model · world-model

Long-horizon real-robot manipulation success rate of 85.0%, process success rate 94.4%, 15 points above the strongest open-source baseline. Existing world action models mostly rely on joint generation, video-first approaches, or decoupled modeling; Riemann-1.0 instead places multi-view observations, robot state, and proprioceptive actions into a single causal sequence, so the same set of weights can both run policy online and serve as a multi-embodiment visual simulator. Training used over 200,000 hours of interaction data, drawn from a mix of first-person human video, handheld gripper demonstrations, and heterogeneous robot trajectories, following a progressive embodiment pretraining approach. In simulation: 94.3% on RoboTwin2.0, 99.0% on LIBERO, and 62.6% on the long-horizon composite benchmark RoboCasa-365.

Haofeng Sun et al. · arXiv 2608.27033 source

WALL-SS: "next-scale" autoregression extends world model usability to minute-scale horizons · world-model

World models have mostly been stuck at clip-level future prediction, drifting once switched to continuous interaction. WALL-SS writes embodied trajectories as a causal sequence interleaving observations and actions, generating each future frame coarse-to-fine across scales, paired with scale-compressed long-range memory: recent interactions keep fine resolution while distant observations and actions are compressed, keeping memory bounded. The authors report coherent streaming rollouts of over one minute, with on-policy alignment used to suppress action drift and long-horizon inconsistency.

Maeve Zhang et al. · arXiv 2608.26239 source

CLAP: cross-embodiment video world model doubles directly as a zero-shot physics simulator · world-model

Action-conditioned video models are usually locked to a single embodiment, leaving internet-scale human video unusable. CLAP reconciles different action spaces using three representations — end-effector pose, language instructions, and latent actions — first learning physical priors on unlabeled video, then grounding to end-effector action space for zero-shot deployment. On DROID it approaches or exceeds single-embodiment SOTA, covering forms including Bridge, dual-arm YAM, and the G1 humanoid, with code and models fully open-sourced.

Kechen Liu, Ola Shorinwa · arXiv 2608.27406 source

SOLO: a humanoid completes a continuous 1.5 km outdoor route using only a single chest-mounted depth camera · locomotion

Perception-based humanoids tend to fail over long distances, for two compounding reasons: dense terrain reconstruction smooths away critical detail, and point-wise imitation lacks temporal credit assignment. SOLO uses Fourier-encoded grid queries to extract evidence from depth and proprioceptive tokens, cutting elevation-map L1 error to a quarter to a third of the baseline; it also writes teacher-student next-state divergence into the PPO reward. In stress-test terrain, average pass rate was 97.5%, 96% on stepping stones, versus 0–3% for the dense-reconstruction version.

Pihai Sun et al. · arXiv 2608.26583 source

PredVLA: a 0.68M-parameter predictive-coding policy outperforms a same-scale transformer by sevenfold · vla

Large models dominate manipulation leaderboards, but it's unclear how much scale language-conditioned control actually needs. PredVLA has only 0.68M trainable parameters, uses no robot-data pretraining at all, and relies on hierarchical generative recurrent dynamics to predict visual features and proprioception, with observations only affecting the hidden state through online inference of prediction error. It averages 86.9% across three short-horizon LIBERO suites, 75.4% including the long-horizon suite. Under identical frontend, demonstrations, and evaluation protocol, it beats a parameter-matched Transformer policy by 3.7x and an LSTM by 7.4x.

Hiroki Sawada, Shunichi Kasahara (Sony CSL) · arXiv 2608.26673 source

Instruct-to-Act: a VLM issues low-frequency instructions while a world-model controller takes over at high frequency · vla

Instruction-tuned VLMs are good at translating instructions into high-level plans but poor at low-latency action sequences in unfamiliar environments; world-model controllers are the opposite — fast but lacking open-ended task guidance. This work relabels controller-policy rollout segments with synthetic instructions, optimizing a behavior-cloning objective jointly with the original reward and world-modeling objectives. Three of the seven embodied environments are multi-agent scenarios where VLMs coordinate via language and a trained controller acts as the executor; swapping in a different pretrained VLM planner requires no fine-tuning.

Zineng Tang et al. · arXiv 2608.26788 source

Learning five juggling patterns with three balls on a real robot in five minutes · manipulation

Models remain useful even when the sim-to-real gap is large. This online learning framework learns local models from accumulated experience while retaining a global prior model to extrapolate in experience-sparse regions, then constructs a mutually reachable set that guarantees transitions between catches never push the arm past joint or actuator limits. A dual-arm robot with multi-fingered hands and onboard vision learned five three-ball patterns — cascade, tennis, half-shower, shower, and box — in under 5 minutes of real-world interaction. Authors include Christopher G. Atkeson and Alfred A. Rizzi.

Taeyoon Lee et al. · arXiv 2608.26800 source

TrapVLA: backdoors that don't just make a robot fail, but dictate how it fails · vla

Prior backdoor attacks on VLAs treated any task failure as a success; this paper requires the attacker to control the failure mode itself — for instance, making a grasp miss by a specified distance — which raises the bar for detection. The authors build a data engine for synthesizing target trajectories along with an automated failure-fidelity evaluation, construct two benchmarks — Trap-LIBERO and Trap-RoboTwin — covering four failure-mode categories, and design a method that explicitly learns trigger-induced action residuals. In both simulation and real-robot experiments, performance on clean data is largely preserved.

Jun-Hui Liu et al. · arXiv 2608.26578 source

Other papers today: FlashVLA uses streaming action decoding to cut inference latency for flow-matching VLAs (arXiv 2608.27384 source); FLARE adds failure detection and autonomous recovery to VLAs (arXiv 2608.26645 source); TemporalFlow-VLA physicalizes execution history to resolve ambiguity between visually similar states in multi-stage manipulation (arXiv 2608.26821 source); Memory Anchors reduces catastrophic forgetting in continually-learning robot policies (arXiv 2608.26545 source); 4DSynth procedurally synthesizes interactive, time-varying embodied simulation scenes (arXiv 2608.26947 source); SpatialCrafter builds walkable worlds from a single image using a generative 3D agent (arXiv 2608.27073 source); RTNav achieves real-time zero-shot object navigation (arXiv 2608.26496 source); MeshPriorDiT hierarchically models action-conditioned cloth dynamics (arXiv 2608.26766 source); Generative Semantic Scene Completion recasts outdoor LiDAR semantic completion as discrete diffusion (arXiv 2608.26737 source).

Open Source · Tools · Benchmarks

· Hugging Face Microduck: a $399 (~¥2,681) open-source bipedal duck-shaped robot, 25cm tall, 770g, 15 degrees of freedom, with a Rockchip RK3566 chip and a monocular camera, plus an open-source SDK, virtual training environment, and reinforcement-learning scripts. Thomas Wolf told Bloomberg that pre-orders were selling roughly one unit every 4 seconds, with sales of about $500,000 and a target of over 20,000 units; the first batch ships before Christmas, manufactured by Seeed Studio (Chinese hardware maker) in Shenzhen source

· Amap ABot-Recon: Alibaba's Amap released a real-time 3D reconstruction solution, claiming to reconstruct scenes with tens of thousands of frames' worth of detail from just 12 input frames source

Funding & Deals

a16z "American Dynamism" fund ｜ New fund ｜ $1.1 billion · adjacent

The money targets the entire physical foundation of AI — chips, memory, interconnects, data centers, robotics — with a16z describing it as "opening the throttle to accelerate AI's physical buildout." A firm known for backing software-scale returns is putting a new fund behind hardware, and the fund's announcement named several gaps: faster and more efficient systems, cheaper and higher-bandwidth memory hierarchies, more scalable interconnects between nodes, and low-power edge devices that let AI explore and interact with the physical world.Source: TechCrunch source

Sharpa ｜ Cumulative funding ｜ over ¥4.5 billion · embodied ⚠️ vendor claim

Industrial investors including Alibaba, Meituan, Tencent, JD.com, and Transsion, alongside institutional investors such as Sequoia China, Qiming Venture Partners, Meituan Longzhu, and Photosynthesis Capital, make up the shareholder list of this dexterous-manipulation company founded in 2024. The figure comes from the company's own disclosure, without a breakdown by round or timing. Sharpa was founded by three co-founders of Hesai Technology (Chinese lidar maker) striking out on their own; co-founder Li Yifan is simultaneously Hesai's CEO. Its flagship offerings are the Sharpa Wave dexterous hand and a companion full-stack robot "brain."Source: Sina Finance source

Mech-Mind ｜ Hong Kong IPO offering ｜ Priced at HK$95.30–101.70 ｜ up to HK$2.706 billion raised · industrial

After the book opened on August 24, the cornerstone investor lineup has emerged: Baillie Gifford leading a group of nine, alongside Taikang Life, Jane Street, BYD's investment arm Golden Link, Invus, Ghisallo, Ernst & Young, NGS Super Fund, and E Fund. Approximately 23.14 million shares are being offered. The company builds the 3D-vision "eyes," AI-decision "brain," and execution "hands" for robots; per a CIC report, it held roughly 22.1% of the global market for AI+3D-vision-guided non-dedicated intelligent robot components in 2025, corresponding to about ¥389 million in revenue — nearly 2.6 times the second-place player's 8.5% share.Source: ACN Newswire source

InstAdapt ｜ Seed round ｜ tens of millions of dollars · world-model

Co-invested by Xiechuang Zhihui, Yunhui Capital, Pudong Science and Technology Investment, and Wuyue Angel, with Xinpai Capital serving as financing advisor. The company was incubated from a ShanghaiTech University team; founder Shi Ye is an assistant professor at ShanghaiTech's School of Information Science and Technology and head of YesAI Lab. Its focus is embodied world-model infrastructure — letting robots switch environments, tasks, or even embodiments without collecting data and training from scratch. Its technical approach was invited to be presented at NVIDIA GTC 2026.Source: PEdaily source

Psi Robot ｜ New round ｜ over $100 million (~¥672 million) · embodied

Following yesterday's disclosure of the round, today the investor list was made public: industrial investors including Tuopu Group, Chery Holding's Ruicheng Fund, and Lens Technology (Chinese electronics component maker), alongside institutional investors 37 Interactive Entertainment, Wuhu Municipal Investment Holding Group, and Fosun Chuangfu, with existing investor Huajin Capital adding an over-allotment. The company's prior angel and Pre-A rounds together totaled ¥2 billion. This funding is primarily aimed at world models: the company's self-developed Psi-R2 manipulation policy model handles long-horizon task decomposition and continuous action planning, while its Psi-W0 action-conditioned world model handles trajectory prediction and correction.Source: cnbeta source

Lingzhang Technology ｜ Angel round ｜ tens of millions of yuan · embodied

The funds will go toward industrializing its dexterous-hand technology.Source: PEdaily source

Xingji Miyuan ｜ Seed round ｜ tens of millions of yuan · embodied

Led by Vertex Capital China, with Woan Robotics and Xuyuan Capital participating. The company positions itself as a supplier of embodied-intelligence systems for all-domain scenarios.Source: Jiwei.com source

Commercialization & Deployment

UBTECH sells 16,123 humanoid units in H1, full-size model revenue up 14x · humanoid

Total revenue of ¥1.27 billion, up 104.2% year-on-year. Total humanoid robot sales of 16,123 units, up 268.3% year-on-year; of which full-size embodied-intelligence humanoid robot revenue was ¥590 million, up 1,445.0% year-on-year. Overall gross profit was ¥570 million, gross margin 44.7%, up 9.7 percentage points year-on-year; the expense ratio fell to 54.4%. The company is still losing money, but the losses are narrowing: adjusted EBITDA was -¥170 million, a 45.9% reduction in loss year-on-year. R&D spending exceeded ¥300 million, with 1,103 R&D staff. The sales figure includes models beyond just full-size units; the company reported the two figures separately.Source: Sina Finance source

DQ's Shanghai Wujiang Road store hands Blizzard-making over to a robot, 55 steps with no human operation · embodied ⚠️ vendor claim

This DQ store in Shanghai, opening August 29, is a collaboration between Sharpa and CFB Group; without altering equipment, ingredients, or process, the robot uses cups, spoons, and blenders designed for human use to complete all 55 steps of making a Blizzard in one continuous run, with Sharpa claiming tactile sensing is directly involved in 98% of them. The above process and figures come entirely from the company's press release, with no independent record yet of sustained operation; the store's long-term reliability will only be known once it has actually been running for a while.Source: Sina Finance source

ICE plans to spend $1–2 million on Boston Dynamics robot dogs · adjacent

A U.S. Immigration and Customs Enforcement notice within the Department of Homeland Security states it will procure Spot robots and accessories for "inspection, situational awareness, and hazard assessment," citing improved officer safety. Spot starts at around $75,000, implying at least several units under the stated budget range. The same day's procurement records also show ICE ordering 6,000 pairs of electroshock gloves for $16.7 million. A Boston Dynamics spokesperson told 404 Media that the company sells to many government agencies and public-safety organizations for uses including hazardous-gas detection, unexploded-ordnance inspection, suspicious-package investigation, and search and rescue, and stressed that any attempt at weaponization is explicitly prohibited under its terms. The New York Police Department previously canceled a similar contract following public backlash.Source: 404 Media source

Thousands of L4 logistics robots from Chengdu's Yangshi Technology now running regular delivery routes on European streets · autonomy ⚠️ vendor claim

These robots are already handling commercial orders for food delivery, supermarket services, and last-mile logistics transfer on public roads in several overseas cities. Each unit has over 60 liters of cargo capacity, a 20 kg payload, supports quick battery swaps, and offers over 12 hours of range per charge; it has 6-wheel drive with independent suspension, can clear obstacles up to 14 cm, and reaches a top speed of 10 km/h. The company plans to deploy 20,000 units nationwide in a single overseas country by the end of 2027 — this is a target, not confirmed orders. Yunda Technology disclosed on an investor-relations platform that Yangshi Technology is its affiliate, in which it holds a 36.7% stake.Source: Guandian.cn source

Japan's Muse delivery robots enter a supermarket in Queens, New York · industrial

Armo One went into service this month at Foodcellar Market, carrying up to 100 kg of goods; staff use an app to summon a loaded robot from the back storage room to the shelves, and after swapping modules it can also scan for out-of-stock items and check price tags. It's billed on a subscription basis, as low as $1 to $2 per hour per unit. Muse's U.S. business head, Nobuaki Momoi, said "about 20% of restocking time goes to moving things around, like pushing carts." Store owner Metin Mangut has been unable to fully staff the store since the pandemic, with some positions going unfilled despite postings; he said, "we believe technology should support people, not replace them." Most U.S. automation investment is concentrated in less-visible places like e-commerce warehouses; retail robots that customers actually encounter remain rare. MarketsandMarkets projects the U.S. physical AI market growing from roughly $300 million in 2026 to about $2.7 billion by 2031.Source: Nikkei Asia source

Industry Developments

Anthropic releases MHS hardware standard, letting AI agents directly drive lab instruments and production-line equipment · adjacent

The Model Hardware Standard is aimed at any device with a programmable interface, covering both scientific instruments and advanced manufacturing equipment; Anthropic compares it to USB-C — a unified way for devices to exchange information. It is not tied to Claude and is model-agnostic; the company says the goal is to shrink the time it takes enterprises to connect and integrate hardware. It's currently in research preview, open only to a handful of organizations in science, robotics, and manufacturing, with plans to open-source it later so that hardware makers in any industry can adopt it. Elizabeth Kelly, who leads beneficial deployments at Anthropic, told CNBC, "We built this for science, to show what AI's potential looks like, but there's enormous value here for enterprise and industry too." The company open-sourced MCP in 2024, which has since become the de facto standard for agents connecting to data sources; it is also building a silicon team for its own chips, recently hiring Caitlin Kalinowski, who previously worked on hardware at OpenAI, Meta, and Apple.Source: CNBC source

China's NDRC: robotics industry must fit local conditions, prevent blind bandwagon-jumping and herd behavior · embodied

Li Chao, deputy director of the National Development and Reform Commission's Policy Research Office, said at an August regular press briefing that the robotics industry touches on numerous frontier technologies including AI, advanced manufacturing, and new materials, and that each region should find its positioning based on local resource endowments and industrial strengths. The next focus areas are embodied-intelligence training grounds and application pilot bases, letting robots iterate on technology in real-world settings. On data, the goal is to build a high-quality real-robot data collection system to address training-data scarcity; on models, the plan is to support companies pursuing multiple technical paths and encourage innovation across vision, language, action, and world models to accelerate technical convergence. On applications, the goal is to expand the library of embodied training scenarios, encourage companies to open up real-world scenarios, and support reliability and safety testing, extending into manufacturing, healthcare, consumer, service, and public-safety sectors. This stance stands in noticeable contrast with the wave of robotics industry plans that various localities have issued in recent weeks.Source: Cailian Press source

Unitree's G1 EDU exposed with two independent root-RCE chains, one starting from Bluetooth · humanoid

Security researcher Olivier Laflamme disclosed CVE-2026-76639 and CVE-2026-76640 on August 27. The first exploits a path-traversal condition in chat_go to reach bashrunner, gaining root on the Locomotion PC. The second starts from BLE proximity: the initial write path doesn't require Bluetooth pairing, and Unitree's cloud service at the time only verified that the requester was a valid Unitree account, not that the account owned the target robot — allowing an attacker to retrieve key material for another G1 EDU, use it to establish an authenticated BLE state, and enter the Wi-Fi provisioning code, where a buffer overflow also leads to root. Unitree patched the cloud-side ownership check in July 2026, which Laflamme says broke his original verification workflow. The exact patched firmware version could not be determined, leaving G1 EDU users without a clear upgrade target; whether the two vulnerabilities affect other models is also unconfirmed. Propagation testing was limited to two robots in the same room.Source: The Hacker News source

Rokae turns profitable in H1, embodied-intelligence business now a third of revenue · industrial

Revenue of ¥415 million, up 136% year-on-year, adjusted net profit of ¥18.4 million, with operating cash flow turning positive. Total robot sales exceeded 13,000 units, up 225% year-on-year. Broken down: industrial robot revenue up 69%, flexible collaborative robot revenue up 72%, and embodied-intelligence robot revenue jumping to ¥139 million — 34% of total revenue — with sales exceeding 5,600 units. Management attributed this to multiple embodied-intelligence companies adopting or switching to its humanoid robotic arms for R&D and product iteration. Gross margin was 29.8%, up 7.5 percentage points year-on-year; R&D spending was about ¥63.9 million, up 146.6% year-on-year.Source: Sina Finance source

Hanoi approves 24-month autonomous-driving pilot, driverless buses, robotaxis, and delivery robots all hitting the road together · autonomy

A decision by the Hanoi People's Committee on August 25 designated two zones. Within the Hoa Lac High-Tech Park, an approximately 11.5 km route will run robotaxis and delivery robots for on-demand passenger transport and delivery; within the Hoa Lac campus, an approximately 2.8 km route will run cleaning and delivery robots, alongside driverless buses ferrying students. Operations run daily from 9:00 to 16:30, with a speed cap of 40 km/h, and are suspended during heavy rain, thunderstorms, flooding, strong winds, or dense fog. Driverless buses and robotaxis must be staffed with a safety operator able to take over at any time, with passenger limits of 8 and 4 respectively. Commercial fares are prohibited during the pilot. The city will temporarily waive registration, licensing, inspection, and operating-permit requirements within the pilot zones, replacing them with alternative safety measures.Source: VnExpress International source

Waymo launches $29.99/month membership, turning robotaxis into a subscription · autonomy

Waymo Premier is invitation-only; its core benefit is 10% back in Waymo Cash per ride, with higher cashback during busy periods. Other perks include priority pickup, priority access when new cities launch, and up to 5 free cancellations per month. Waymo says this was requested by users. At $29.99/month, the price implies riders need to use it frequently enough to break even.Source: Mashable source

Hardware · Supply Chain

· LinkTech (Chinese harmonic reducer maker): harmonic reducer capacity has reached 1 million units, targeting 1.5 million by year-end, with over 60% market share in China; order visibility extends to Q3 2027, with humanoid-related orders making up over 40% of the total; its weight-reduced version for humanoids is 20–45% lighter than the standard version. On August 26 its board approved a proposal to issue H-shares and list in Hong Kong source

· Yeoyo Technology: joint orders on hand for 2026 have surpassed 1.3 million units, with full-year deliveries projected at 1.3 million — versus just 95,000 units shipped in all of 2025; its completed automated joint production line has an automation rate of 85%, with a 60-second cycle time per harmonic joint source

· Frameless torque motors: Leadshine Technology's on-hand orders for humanoid frameless torque motors exceeded 1 million units as of H1, with a production line targeting an annual capacity of 3 million units also being built out; Lingqiao Drive & Control's on-hand orders for frameless motors passed 1 million units in H1, versus just 120,000 for all of 2025 — its general manager said, "the biggest pressure right now is delivery" source

· Six-axis force sensors: Kunwei Technology has annual capacity of 60,000 units, targeting 120,000 by year-end, already supplying UBTECH, AgiBot, Galbot, and Xiaomi at scale; Xinjingcheng's 6.5mm- and 9.5mm-diameter micro six-axis force sensors are already deployed at scale in dexterous-hand projects, with Q1 shipments exceeding 100,000 units and a full-year target of over 500,000 source

· Gap between planned capacity and actual delivery: the year-end capacity figures and full-year delivery targets above are all public plans, not yet realized; LinkTech's own capacity-expansion project via a follow-on share offering is only 8.39% complete, with completion pushed back two years source

· Planetary gear reducers: DigiTimes reports that humanoid robot demand is driving up shipments of planetary gear reducers source

· Nanshan Zhishang: its self-developed high-strength transmission tendon rope achieved small-batch orders in H1 source