Key Points Today

· Kodiak signs its first driverless heavy-truck supply contract: 75 trucks go to DTL from 2028

· Waymo adds nearly 250 square miles of driverless service area across 5 cities at once

· Unitree (Chinese humanoid robot maker) posts its first quarterly report since listing: Q3 revenue up 69.93%, net profit excluding non-recurring items down 17.42%

· North American Q3 robot orders reach 10,989 units, with automotive component orders up 90.4%

· Rollback rewinds a world model to the decision point and retries; after 8 hours of training, it beats the average human player's success rate on two real-robot tasks

· Google's Genie 4 simulates the same scene coherently for 2 minutes, twice the previous generation

PDRVLA: Teaching a VLA physical dynamics without generating videovla

HF 134↑, the most popular in the community today. Attaching a video world model to a VLA adds the latency of generating future video to the control loop. PDRVLA instead uses a lightweight physical-dynamics module that extracts "physics tokens" from vision-language features and uses them directly as conditioning for action generation. During training it predicts future RGB, depth, and optical-flow features, with auxiliary supervision from a frozen video world model, and all data comes from public datasets. Across four simulation benchmarks, it reaches a 99.3% success rate on LIBERO and 93.6% on RoboTwin 2.0.

Hanqing Wang et al. · arXiv 2610.37381 source

Rollback: Rewinding the world model to a decision point and trying a different actionworld-model

The real environment can't be rewound, but a world model can. Rollback returns to a key decision point, branches out to try other actions, then compares the outcomes of the branches, attributing success or failure to individual decisions, so no separate value model needs to be trained. Data generated by real-robot execution is then fed back to correct the world model's dynamics via LoRA. On three real-robot tasks, the success rate is 10.67 to 11.85 percentage points higher than the strongest baseline. Training starts from teleoperated demonstrations well below human level, and after 8 hours it surpasses the average human player's success rate on both ball throwing and tabletop curling. HF 50↑.

Jiayi Luo et al. · arXiv 2610.36068 source

ThinkWorld: Human videos without action labels can directly supervise a VLAvla

HF 45↑. ThinkWorld writes each step of "action thinking" as a string of discrete motion tokens, called Chain-of-Dynamics, which describes how the current state changes into a future state, and this representation is shared across embodiments. The model is first initialized with action-labeled demonstrations, then predicts what future each CoD segment leads to, and the segment that better matches the actual observation is reinforced. This lets human video and robot data train the same reasoning space. Both simulation and real-robot experiments show effective human-to-robot transfer.

Chaehyun Kim et al. · arXiv 2610.37666 source

E0.5: A world action model with a 3D foundation model as its backboneworld-model

World action models (WAMs, models that predict both the future and actions) mostly use a video model or a VLM as the backbone, with 3D geometry only as an auxiliary input. Era AI Lab's E0.5 uses a 3D foundation model directly as the backbone and predicts future 3D states and actions together. Without embodied pretraining, its success rate under the clean setting of RoboTwin 2.0 is 89.90%. After pretraining on more than 160,000 trajectories from AgiBotWorld-Beta, average success rates on two real robots, AgileX Piper and Zerith H1, rise by about 15 and 7 percentage points respectively.

Era AI Lab Team et al. · arXiv 2610.36011 source

JAM: The VLA vs. WAM debate, split into three separable choicesvla

The two camps disagree on whether a policy should inherit knowledge from vision-language models or from visual generation models. The authors break the debate into three choices: backbone initialization, the training objective for future observations, and whether to keep this part of the structure at inference time. The resulting JAM adds only a "foresight" loss; the prediction head is discarded after training, and neither training nor inference generates images. Compared with FastWAM, the LIBERO-Plus success rate rises from 55.3% to 64.5%, and estimated FLOPs per sample during online training are 73% lower. HF 3↑.

Yonghao Yu et al. · arXiv 2610.37219 source

OmniBrain-0: Xiaomi hands manipulation and navigation to a single decision modelvla

From the Xiaomi embodied intelligence team. Task understanding, planning, progress tracking, and failure recovery are all handled by one high-level model. Training has two steps: first, interaction experience from strong teacher models is distilled into two experts, one for manipulation and one for navigation; then multi-teacher on-policy distillation merges the two back into one small model. A separate Meta-Agent rewrites execution logs into scoring rubrics and verified trajectories, continuously feeding them back into training. The model was evaluated on five benchmarks, including LIBERO-Pro, RoboCasa365, and R2R-CE, as well as real-robot tasks, with inference and execution costs far below those of its closed-source teachers.

Xiaomi Embodied Intelligence Team et al. · arXiv 2610.36013 source

GAWM: A world action model under 100 million parameters beats billion-scale rivalsmanipulation

GAWM has under 100 million parameters and no robot or video pretraining. It outperforms billion-parameter WAMs on RoboTwin 2.0 and matches billion-scale VLAs on LIBERO. The method predicts the future at two time scales in a frozen DINOv3 latent space, then regresses action chunks directly from the current tokens and the predicted tokens. Ablations show the gains from the latent-space world model are largest when training data or training time is limited. The code is open-sourced.

Yang Liu et al. · arXiv 2610.37091 source

From 12 to 200 demonstrations: A real-robot comparison of four common policies on a low-cost armbenchmark

This single-author paper ran 1,340 scored real-robot episodes on an SO-101 arm, comparing ACT, Diffusion Policy, SmolVLA, and π0.5. With little data, π0.5 is best, succeeding in 16 of 40 trials with 12 demonstrations. At 100 demonstrations, ACT trained from scratch reaches 37 successes out of 40 and overtakes π0.5. For π0.5, the fine-tuning method matters more than data volume: with LoRA fine-tuning, only 4 of 110 episodes succeed; fine-tuning the action expert yields 130 successes in 200 episodes. Two groups of 10 episodes each, run on the same checkpoint, differed by 2 successes on average.

Simon Cowling · arXiv 2610.37170 source

Physical Intelligence ｜ New round (early talks) ｜ At least $1 billion ｜ Valuation about $20 billionworld-model ⚠️ Rumor

Bloomberg, citing people familiar with the matter, says existing investors Lux Capital and Avenir are in talks to lead. Its valuation earlier this year was $11.5 billion, so this round would nearly double it. The company currently has no revenue, and this month opened fine-tuning access to its latest robot model in beta to outside developers. Rival Skild AI was valued at $14 billion in January, and its annual recurring revenue recently surpassed $100 million. If the deal closes at $20 billion, zero-revenue PI would be valued above Skild AI. Source: Bloomberg (relayed by Investing.com) source

Elite Robots (Chinese cobot maker) ｜ Series D++ ｜ RMB 200 million ｜ Valuation undisclosedindustrial

Led by CCB Equity (China Construction Bank–linked investor), with strategic shareholders following. Accounts of the round differ: Sina Finance writes Series D++, while DoNews writes Series D+; Zhidx, citing Qichacha, says the company already completed an RMB 600 million Series D+ in the first half of this year. Founded in 2016, Elite started with controllers and focuses on collaborative robots; it once won a single order of 3,000 cobots from BYD, and its customers also include Luxshare (Chinese electronics manufacturer) and Lens Technology (Chinese glass and electronics maker). The funds will go toward expanding into optical communications and AI-server production-line scenarios, and toward expanding capacity at a new factory. Exports accounted for 33% of 2025 sales volume. Source: Zhidx source; Sina Finance source; DoNews source

Shanyu Chuang Program (Sequoia China × Unitree) ｜ Early-stage investment program ｜ Amount undisclosedadjacent

On the day Unitree released its quarterly report, Sequoia China (venture investor) and Unitree announced they were jointly launching this program to invest in early-stage projects in embodied intelligence, robot applications, and key upstream and downstream technologies, especially young teams. Wang Xingxing (Unitree founder) said Sequoia has been "one of Unitree's earliest and most important supporters" since 2019. Neither the fund size nor the two sides' contribution ratio was disclosed. Source: Sequoia China source

Lige Wanwu (new humanoid foundation-model startup) ｜ First round ("friends and family round") ｜ Amount undisclosed ｜ Rumored valuations of about $70 million and $200 million across two roundsworld-model ⚠️ Rumor

It was previously reported that DeepSeek was investing in a new company of Yu Tao, former head of humanoid foundation models at LimX Dynamics (Chinese humanoid robotics firm); now the company name and its backers have been reported: Lige Wanwu's first round was funded by Sequoia China, Lisi Capital, and DeepSeek. According to Leiphone's Jingxi (Chinese tech outlet), the round was initiated by Liang Wenfeng (DeepSeek founder) personally, who brought in Sequoia, which in turn introduced Lisi. The company has not made an official announcement. Its official positioning is "large language models + embodied intelligence"; Embodied AI Home (Chinese industry outlet), based on Yu Tao's papers, judges that it is building a humanoid behavior foundation model. Source: Embodied AI Home source

Qianxing Innovation (Chinese underwater-robot maker) ｜ Series C+ ｜ Nearly RMB 100 million ｜ Valuation undisclosedadjacent

Shenzhen Guangju Investment, Zhejiang Yunchuang Qizhen, and others participated, just over half a year after the nearly RMB 100 million Series C in March. All three founders come from Huawei. The company makes industrial-grade underwater robots, pool-cleaning robots, and fishing gear, with overseas business consistently above 85% of the total; its Ganzhou base has a designed capacity of 250,000 units per year. Source: Zhidx source

Kodiak signs its first driverless truck supply contract, 75 trucks to DTLautonomy

Starting in 2028, Kodiak will deliver 75 self-driving trucks to California carrier DTL Transport. According to TechCrunch, this is Kodiak's first binding commercial contract for driverless heavy trucks on public highways. In September the two companies had just trial-run fresh produce on California's Highway 99, with a person still in the driver's seat at the time. DTL will first run two lanes: Dallas to El Paso, and Dallas to Atlanta on I-20, both of which Kodiak has already tested and operated on. Kodiak plans to remove the safety driver from the cab in December, starting with hauling for IKEA on I-45 between Dallas and Houston. Founder and CEO Don Burnette said: "We think this is a real inflection point for the entire autonomous driving industry." Existing customers Charger, Roehl, and Werner use a TaaS model, with Kodiak's own fleet hauling on their behalf, and Burnette himself says these operations are still at the pilot stage. Kodiak ultimately wants to shift to DaaS: customers buy and operate the trucks themselves, and Kodiak charges a technology fee. Its stock is already 87% below its opening price on its first trading day. Source: TechCrunch source

Waymo adds nearly 250 square miles of service area across 5 cities in one weekautonomy

Los Angeles, Nashville, Orlando, San Diego, and the San Francisco Bay Area, totaling nearly 250 square miles, all cities where Waymo already operates driverlessly. In Los Angeles, coverage extends south along the coast to Torrance. San Diego expands north to La Jolla, even though the service only opened to the public in that city on September 1. Bay Area service now runs from San Francisco down the Peninsula to San Jose and Campbell, and also connects to BART's Berryessa station. Waymo did not break down the area by city or say how many vehicles it added for this. According to Electrek, it currently operates in 15 US metro areas, with more than 500,000 paid rides per week. In the same week, Robotaxi Tracker counted Tesla's Houston service area as expanded to about 254 square miles, with only 34 vehicles running. Source: Electrek source

Tesla's Robotaxi registrations in Texas rise to 1,080, with Cybercab outnumbering Model Y for the first timeautonomy

On October 27, the Texas DMV registry listed 1,051 authorized vehicles under Tesla Robotaxi, LLC: 544 Cybercabs and 507 Model Ys, the first time Cybercab has been the majority. A few hours later another 29 Cybercabs were added, bringing the total to 1,080. On October 9 the same registry showed 739 vehicles, of which 319 were Cybercabs. The registry lists only VINs and models, so it is impossible to tell which city a vehicle is in or whether it is carrying passengers. Tesla has confirmed that Cybercab production is outpacing deployment, and the line sometimes pauses to avoid a backlog. Source: TeslaNorth source; Drive Tesla Canada source

Otto Group pilots Sereact dual-arm systems for returns handling at a Polish logistics centerembodied

Phase one deploys 10 dual-arm systems at an Otto Group logistics center in Poland to inspect and sort returns; phase two, depending on results, adds 15 more arms, for a total of 25. Only after phase two succeeds will Otto evaluate rolling out across its entire logistics network. Sereact's Cortex model uses NVIDIA Isaac Sim and Isaac Lab to expand real grasps from the returns line into many simulated variants, then uses Cosmos to re-render them with different garment states, packaging, and lighting. The company says that when a robot meets an item it can't grasp well, the corrected policy can be pushed out within days. Sereact was previously handling Zalando returns at a CEVA warehouse and has partnered with Teradyne. At Otto, this is still pilot scale. Source: GlobeNewswire source

Nearly 100 driverless electric mining trucks run in round-the-clock convoys at Huaneng's Yimin open-pit mineautonomy

The Yimin open-pit mine, run by Huaneng (Chinese state-owned power group), is in Hulunbuir, Inner Mongolia, where winter lows can reach minus 40℃, and nearly 100 driverless electric mining trucks run in convoys around the clock. Each truck carries 90 tonnes, is fitted with 5 lidars and 4 millimeter-wave radars, and sends perception data in real time to cloud dispatch 6 kilometers away. The first batch of 100 vehicles maintains an attendance rate above 90%, and daily haulage exceeds 100,000 tonnes. When the charge runs low, a truck drives itself into a battery-swap station, where a robotic arm swaps the battery in 6 minutes. The mine has 12 mobile 5G base stations, and in September it completed 5G-A gigabit uplink verification in an open-pit mining scenario. Source: China News Service source

Unitree's first quarterly report since listing: Q3 revenue up 70%, profit down 17%humanoid

Q3 revenue was RMB 665 million, up 69.93% year on year. Net profit attributable to parent shareholders was RMB 113 million, down 17.37%; net profit attributable to parent shareholders excluding non-recurring items was RMB 105 million, down 17.42%. Single-quarter R&D spending was RMB 127 million, up 249.58% and 19.1% of revenue; the company explained this was mainly due to a sharp increase in R&D and compute spending on large embodied models. First-three-quarter revenue was RMB 1.817 billion, up 55.71%; net profit attributable to parent shareholders was RMB 387 million, up 270.54%, with the surge mainly due to a large one-off share-based payment expense recognized in the same period last year. The South China Morning Post reports that net profit excluding non-recurring items for the first three quarters fell 18.78% year on year, and operating costs rose about 78% to RMB 836 million, growing faster than revenue. In the first half, revenue grew 48.54% and net profit excluding non-recurring items fell 19.34%. Q3 revenue growth was more than 20 percentage points higher, yet profit excluding non-recurring items still declined by double digits. After IPO proceeds arrived, cash and cash equivalents rose from RMB 1.419 billion at the end of last year to RMB 7.354 billion. Ahead of the report, the stock rose 8.74% that day to close at RMB 695.90. Source: Sina Finance source; South China Morning Post source; Reuters source

Google DeepMind releases Genie 4, simulating a single scene coherently for 2 minutesworld-model

Genie 4 is currently open only to trusted testers as a research preview, and will then open in turn to Google AI Ultra subscribers and early partners on the Gemini API and AI Studio. According to TechCrunch, the same scene can stay coherent for up to 2 minutes, twice as long as Genie 3. Visuals are more realistic, and music, ambient sound, and character dialogue are generated in sync. During a simulation, users can use text to change the weather, move the camera, and add people or objects, with Gemini fitting the instructions into the scene. When the user provides no input, Gemini also monitors the scene to keep it advancing coherently. DeepMind says Gemini Omni is responsible for keeping the simulation close to real-world appearance and behavior. In an experiment related to Gemini Robotics, a Gemini model breaks a task into steps and Genie 4 simulates each step in real time, used to rehearse robot scenarios before real-robot execution. Google research director Tejas Kulkarni said: "We'll increasingly see this kind of hybrid world model, with Gemini writing code and text and Genie reading it in real time, simulating reality at different levels of abstraction." Source: Google DeepMind source; TechCrunch; Wall Street CN source

North American Q3 robot orders reach 10,989 units, up 24.8% year on yearindustrial

Data from A3 (the Association for Advancing Automation): North American companies ordered 10,989 robots in Q3, worth $723.5 million. Units rose 24.8% year on year and value rose 26.1%, making it the strongest quarter of the year. The first nine months total 28,984 units and about $1.89 billion, up 9.6% and 13.3% respectively. Automotive component orders rose 90.4%, and semiconductor and electronics/photonics orders rose 97.6%. Automaker (OEM) orders fell in the first half and rebounded 7.5% in Q3, but are still 17% below the same period last year over the first nine months. Food and consumer goods orders fell 23.7%. Source: Robotics Tomorrow source

LG Electronics raises its robot training data target for this year from 100,000 hours to 500,000 hoursadjacent

Lee Jae-mo, head of robotics strategy at LG Electronics, said on the Q3 earnings call that, combining real and virtual data, it aims to accumulate about 500,000 hours of training data within this year. The year-end target the company set in August was 100,000 hours, so this is 5 times higher. The data comes from a robot data factory in Yangjae-dong, Seoul. CFO Kim Chang-tae said the factory formally goes into operation in Q4 and will be Korea's largest robot training and data collection site. LG's cooperation with Nvidia spans simulation, world models, and robot development platforms. Paid external pilots and data sales are pushed to the medium to long term, and Lee said the current priority is "internal use." ⚠️ PlanSource: Chosunbiz source; source

GM: 2028 Escalade IQ gets highway autonomous driving, keeping both radar and lidarautonomy

In an interview, GM chief product officer Sterling Anderson said that with cameras alone, by the most basic laws of physics, recognition will always fail under some conditions; radar is already cheap, so there is no need to remove redundant sensors to save cost. Under GM's plans, the 2028 Cadillac Escalade IQ will carry a new version of Super Cruise that keeps cameras, radar, and lidar all together, and highway autonomous driving will launch by 2028 at the latest. Source: Kuai Keji (Sina Tech) source; Wangtongshe source

South Korea plans to deploy about 1,700 AI robots in the public sector in 2027adjacent

This is a plan from South Korea's Ministry of Trade, Industry and Energy: about 1,700 units to be deployed in 2027, mainly wheeled and quadruped robots, for research, security, policing, and agriculture, with operating data fed back to robot companies. A matchmaking event was held in Seoul on October 28-29, where 9 departments and agencies, including the Ministry of National Defense, the Army Headquarters, the National Fire Agency, and the National Police Agency, discussed needs face to face with 11 companies including Rainbow Robotics, Robotis, Hyundai Motor, and Newbility. Accompanying guidelines will be finalized by year-end. Source: AI Insider source

Tensor signs a letter of intent with Munich taxi center to deploy up to 500 Robocars in Germanyautonomy

Tensor signed a letter of intent with Munich taxi center MTZ: initially up to 500 Robocars to be deployed in German cities, conditional on a formal agreement and L4 approval. The vehicles would be operated by MTZ under its own brand, and passenger data would also belong to MTZ. The two sides also proposed placing 10,000 vehicles with European fleet operators within 24 months of the vehicles entering Europe; the announcement itself notes this is not a firm order. ⚠️ PlanSource: AI Insider source