FutureX · Physical AI Daily — Issue 65 (07/22)

Today's Highlights

· Samsung establishes robotics division RX reporting directly to the CEO, poaches former Boston Dynamics strategy lead

· AGIBOT (Chinese humanoid robotics startup) reportedly pushing for IPO at a $20 billion valuation, with revenue projected at RMB 4 billion this year

· Nvidia open-sources Cosmos 3 Edge: a 4B-parameter world model that generates actions directly on-device on robots

· Momenta secures Robotaxi road-test permit in Shenzhen, stock up over 10% in a single day

· Alibaba releases 122B embodied foundation model RynnBrain 1.1, beating closed-source models across three spatial reasoning leaderboards

I. Research Progress

RynnBrain 1.1: 122B embodied foundation model beats closed-source models across three spatial reasoning leaderboards · vla

Alibaba DAMO Academy has pushed its open-source embodied foundation model to 122B-A10B scale, currently the largest open-weight tier publicly available, with HF trending at 30↑. Compared to version 1.0, the full lineup adds contact-point prediction, and the 2B/9B variants gain native 3D grounding, making outputs more tightly aligned with actual mechanical manipulation. The accompanying RynnBrain-VLA uses a unified cross-embodiment action space plus embodiment-specific masks, and has already been deployed on three robot platforms: Unitree G1, Astribot-S1, and Tianji Wuji. The 122B-A10B model surpasses all closed- and open-source models evaluated in the paper on VSI-Bench, MMSI, and RefSpatial-Bench; in real-robot experiments, policies initialized with it also outperform the Qwen series and mainstream general-purpose VLAs.

Kehan Li et al. (Alibaba DAMO Academy) · arXiv 2607.17977 · Analysis: Embodied Intelligence and Spatial Perception

POT-VLA: Giving humanoids memory that "the object is still there," success rate up from 39/80 to 71/80 · locomotion

A hidden failure mode in long-horizon humanoid loco-manipulation is "object state divergence": the object state used to drive whole-body actions and the state used to judge whether an action achieved its goal are actually two separate representations. The authors use persistent object tokens to maintain role-indexed 3D object records from RGB-D input, with the same record both conditioning action generation and supporting geometric predicate verification, closing the execution loop. Across eight real-world task categories on Unitree G1, the aligned GR00T-N1.7 baseline improved from 39/80 to 71/80.

Peng Ren et al. · arXiv 2607.18016 · Analysis: AI Era (Chinese AI media outlet)

Adding a depth prior to JEPA world models lets an 18M-parameter model handle in-the-wild data · world-model

Complex, unpredictable outdoor real-world robot video has long been a challenge for JEPA-style world models. The authors introduce depth supervision plus isotropic latent-space regularization during training, over-parameterizing only at training time with no added inference cost. Training an 18M-parameter model on agricultural robot video, the frozen representation's error on a visual odometry probe is 33% lower than the LeWM baseline, and its "surprise" discrimination on out-of-domain TartanGround also clearly widens, with the advantage growing with rollout length.

Usman M. Khan · arXiv 2607.16314

MEVION: A $14,000 open-source dual-arm data collection rig that fills ALOHA's gap in handling heavy loads · benchmark

ALOHA is already the de facto standard for dual-arm data collection, but its force and speed ceilings are low—heavy objects and fast motions are both off the table. The authors use sheet-metal welding to build a large-frame body from a small number of parts, all sourceable off-the-shelf; the four 6-DOF arms weigh 7.0 kg each with a peak torque of 60 N·m, and the full setup costs around $14,000. The elbow joint borrows a closed-chain mechanism from quadruped robots to reduce distal mass. Both hardware and software are fully open-sourced.

Kento Kawaharazuka et al. · arXiv 2607.17970 ); PhyAgentOS provides a unified runtime for heterogeneous embodied agents, separating "execution termination" from "semantic completion" (arXiv 2607.16636 ); FM-VLA adds force-sense memory to VLA for contact-rich manipulation (arXiv 2607.18231 ); an asynchronous multimodal diffusion policy handles the mismatch between each modality's perception rate and inference latency (arXiv 2607.17257 ); A2RL Vmax provides a long-distance, high-speed perception and multi-vehicle interaction dataset for autonomous racing (arXiv 2607.17813 ); UniETP unifies observation formats and action types across different simulators for general-purpose embodied task planning (arXiv 2607.18062 ); importance sampling plus PCA finds rare failures in a commercial autonomous trucking stack (arXiv 2607.18106

· lingbot-world-2: Claimed to be the world's first hour-scale real-time world model, now open-sourced—moving the camera generates responsive scenes in real time · The Standard · Robo Encyclopedia

Striverobotics ｜ Series C ｜ Hundreds of millions of RMB · industrial

Led jointly by Zhejiang state-owned Qianggang Capital and global optical module leader Innolight, with Guangtou Capital, Guohai Securities, and Juntong Capital following, and existing shareholder Fosun RZ Capital increasing its stake. The company builds embodied robots for high-altitude specialized work—a niche overshadowed by the humanoid boom but with among the most rigid demand for replacing dangerous human labor.Source: Tech Capital Circle

Sereact (Germany) ｜ Series B extended to $116 million ｜ Zalando takes strategic stake · industrial

European e-commerce giant Zalando has joined this Berlin-based warehouse robotics company's Series B as a strategic investor, with Headline leading; total funding raised now exceeds $145 million. More than 200 Sereact systems are currently running across Europe, with customers including Mercedes-Benz Trucks and BOL. The funding will primarily go toward its next-generation robot brain, Cortex 2.0, whose approach is to use a learned world model to predict physical outcomes in advance and score candidate trajectories by risk and efficiency—shifting warehouse automation from "try it and see" to "compute first, then act."Source: Retail Technology Innovation Hub

Industry data: Global robotics startup funding hits $18.8 billion in H1, already surpassing all of 2025 · adjacent

A Crunchbase report shows global robotics startup funding reached $18.8 billion (approx. RMB 136 billion) in H1 2026, far exceeding all of 2025's $15 billion—a historic high. The trend is even steeper in China: by IT Juzi's count, embodied intelligence funding in China reached RMB 93.5 billion in H1, roughly 5x year-over-year, with 322 deals, up 137% year-over-year, and the number of unicorns valued above RMB 10 billion rising from 3 in 2025 to at least 26. In Hong Kong, 10 companies with robotics businesses have already listed in H1 this year with 16 more in the pipeline—yet the two that listed earliest, Huayan Robotics and Ledong Robotics, have both seen their stock prices fall by nearly half from their highs, suggesting secondary-market price recalibration has already begun.Source: Medical Device Business Review (Chinese medtech outlet)

III. Commercialization and Deployment

Tesla expands Robotaxi to Tampa and Orlando; Cybercab to feature built-in Starlink antenna · autonomy

Following its expansion to Miami this month, Tesla has pushed Robotaxi service to two new Florida cities, Tampa and Orlando, ahead of its Q2 earnings report. Also announced the same day: Cybercab will integrate a Starlink V5 antenna—for a driverless fleet, remote takeover and telemetry links in cellular dead zones have long been an operational hard constraint, and satellite backhaul fills that gap. Robotaxi and Optimus progress are also expected to be the focus of Wednesday's earnings call, with markets watching whether cash burn can translate into a per-vehicle profitability model.Source: Reuters

Momenta Robotaxi secures road-test permit in Shenzhen, stock up over 10% in a single day · autonomy

Shenzhen becomes Momenta Robotaxi's fourth core city in China after Shanghai, Suzhou, and Wuxi, with overseas testing already underway in Munich, Germany and Abu Dhabi, UAE. The market reaction was immediate, with the company's stock up over 10% that day. A previously reported mass-production custom vehicle project with SAIC's Xiang Dao is still on track for a 2027 reveal, with SAIC manufacturing, Momenta providing the L4 solution, and Xiang Dao operating—a division of labor that shifts Momenta from a Tier 2 solution seller toward a deep partnership with the automaker.Source: Eastmoney

BMW locates humanoid robot software center in Landshut, prioritizing data and simulation before production-line deployment · industrial

BMW has designated its Landshut, Germany components plant as the central development site for humanoid robot software, handling training data generation and preparation, virtual simulation, motion planning, and robot training—splitting responsibilities from existing production-line pilots in Leipzig and Spartanburg, which validate specific applications while Landshut builds the technical foundation. The technical approach is an open modular architecture combining fixed programmed workflows with AI models, where a VLA model translates visual information and task descriptions into concrete actions; data collection uses cameras, motion-capture suits, and data gloves. Project lead Christoph Jagoda emphasized that the evaluation covers not just technical feasibility but which processes humanoid robots can genuinely deliver economic returns for. Partners include Athenyx Robotics and Landshut University of Applied Sciences.Source: Automotive Manufacturing Solutions

Incheon Coupang warehouse fire burns for 61 hours, hundreds of lithium-battery robots on the fifth floor a hazard concern · industrial

A fire at a Coupang logistics center in Incheon, South Korea took 61 hours to fully extinguish, possibly the longest on record in the area, with flames burning through six floors; hundreds of lithium-battery-powered robots on the fifth floor were at one point flagged as an explosion risk. On-site investigation points to melted roller shutter doors and extremely high fire load. This highlights an under-discussed boundary condition as warehouse automation scales up: densely packed lithium-battery-powered robots are themselves a concentrated fire load, and existing fire-safety designs may not have been built around that premise.Source: Seoul Economic Daily

RoboSense forms alliance with Abu Dhabi's Origen, pushing embodied intelligence into the Middle East · autonomy ⚠️ letter-of-intent stage

RoboSense (HKEX: 2498) signed an agreement with Abu Dhabi AI company Origen during WAIC, covering spatial intelligence, autonomous driving, industrial inspection, and smart cities, with Origen handling regional delivery, solution integration, and localization. Both sides currently describe the arrangement as "evaluating pilot deployment and localization strategy," with no specific orders or installation figures yet.Source: Zawya · Chosunbiz · GIGAZINE

Beijing's industrial robot output up 75.5%, service robot output up 2.3x in H1 · industrial

According to Beijing Municipal Bureau of Economy and Information Technology data: in H1, "AI+" initiatives drove output growth of 92.1%, 75.5%, and 2.3x respectively in integrated circuits, industrial robots, and service robots. Over the same period, the city's combined industrial and information software sector value-add approached RMB 1 trillion, or 36.6% of GDP, with fixed-asset investment in key industrial sectors up 24.0% year-over-year, including a 36.0% increase in high-tech manufacturing investment. The gap between service robots' more-than-doubled growth rate and industrial robots' rate suggests this wave of volume growth is landing more in commercial service applications than on production lines themselves.Source: Beijing News · Kanchai.com (Chinese tech news outlet)

XPeng unveils TuringViT vision encoder · autonomy ⚠️ manufacturer's own claims

XPeng has unveiled a vision encoder for physical AI applications called TuringViT, claiming it restructures the training paradigm for large vision models. Only the launch announcement is public so far, with no third-party reproduction or benchmark comparison yet seen.Source: Gasgoo (Chinese automotive industry outlet)

· Data gloves · a collection-side bottleneck: mainstream Manus gloves cost RMB 60,000–100,000 per set and prices are still rising while supply remains scarce; pairing with an RMB 100,000-class dexterous hand pushes per-set collection cost toward RMB 200,000–300,000, prompting BrainCo (a 21-DOF exoskeleton, full set RMB 160,000), Dexrobot, and Linjiedian to each build their own gloves

· Weihong Corporation · micro servo motors: claims its dexterous hand's micro servo motors overcome the stalling and burnout issues of coreless motors, with projected monthly output of 80,000 units by year-end ⚠️ manufacturer's own claims