The University of Hong Kong and Tsinghua University propose VAST (Horizon Adaptive Offline Policy Learning via Value Stitching), an offline value learning framework that adaptively selects update horizons by jointly choosing future subgoals and time spans to stitch local returns with subsequent values.
Limpid Intelligence founder Li Yiming's team at Tsinghua University released the world action model PrismWAM and the execution diagnosis framework RoboDoctor. PrismWAM decouples semantic context from physical state and, without additional action-data pretraining, achieved an average success rate of 83.67% on RoboTwin 2.0 and 75.5% in the Clean-to-Random setting, ranking first on the official leaderboard.
At the ThunderWorld 2026 technology conference, ThunderSoft released AquaDrive AIOS 3.0, upgrading around three capabilities: autonomous generation, autonomous fusion, and autonomous control. The autonomous fusion-related technology has been retrofitted onto a Chery Tiggo 8 production platform and validated on real urban roads.
On October 13, Physical Intelligence (PI) announced the launch of the Pi API beta, packaging robot data upload, hosted fine-tuning, cloud inference and policy generation, and evaluation feedback into a single service chain, in an attempt to sell embodied brains via API.
In the first seven months of 2026, RMB 66.6 billion flowed into the world model sector in China's primary market, with 58 companies securing funding, but by strict pretraining criteria, no more than five are truly doing it. Overseas institutions are competing for talent with high salaries: 1X is hiring for world model roles at USD 200,000–300,000 per year plus equity, and Runway, xAI and others have also set up positions. Autonomous driving, humanoid robots and physical AI, gaming 3D, and industrial digital twins are the main application directions, and the sim-to-real gap remains the top challenge.
Silicon Valley 101 invited Pokee.ai founder Zhu Zheqing, Pokebot founder Xu Huazhe, and University of Washington assistant professor Wang Sheng to discuss the impact of GPT-6 Astra's release on three paths: personal agents, robotics, and AI drug discovery.
RoboJev-27B is an open-weight vision-language model that takes a task instruction and 10 frames of a manipulation attempt to judge whether the robot completed the task. It achieves 95.2% balanced accuracy and 0.993 AUROC on 268 held-out episodes of the same robot model (SO-101), and an AUROC of 0.890 on 459 episodes from unseen labs and robots. Median inference time is 0.26 seconds per episode on a single H100, and the model, inference code, and evaluation results for 29 model/protocol entries are released together.
At Tesla's Q3 2026 earnings call, Musk revealed that the company is building a robot training system called Optimus Academy and plans to manufacture about 10,000 to 20,000 Optimus robots so they can generate training data through autonomous practice.
A team led by Xie Guangming and Zhang Feitian at Peking University proposed a hybrid model that uses an equilibrium anchoring model and a lightweight correction model to suppress the accumulation of step-by-step prediction errors in flexible robotic arms. The work was published at RSS 2026. When predicting 200 steps (6 seconds) on a simulated 3D cable-driven robotic arm, the hybrid model's error was more than a quarter lower than that of the anchoring model alone, and none of 19 tests showed loss of accuracy. On a physical hydraulic soft robotic fish tail, its error was half or less that of the anchoring model and a quarter or less that of the correction model. More stable predictions could let a controller first try out multiple actions in a computer and pick the safest one.
Generative Bionics announced this week plans to build a multi-site Physical AI Data Factory in Italy, with the first hub covering 5,000 square meters and expected to begin operations in the fourth quarter of 2027.
Dexory announced a collaboration with NVIDIA to advance physical AI in warehousing: DexoryView Storage Health integrates NVIDIA Cosmos, using foundation model reasoning to identify storage anomalies such as opened boxes, misaligned pallets, and misplaced goods.
At the "2026 World Physical AI Conference and Partner Ecosystem Conference" hosted by Unitree Robotics, Wang Xingxing said the ChatGPT moment for embodied intelligence could arrive as early as 2027 or 2028, and proposed a "double 80% standard": completing 80% of tasks in 80% of unfamiliar scenarios using only voice or text instructions.
DEEP Robotics has released five research results centered on its self-developed embodied brain, DEEPRobotics-Brain, including the world model DR-WM-1.0, the embodied foundation model DR-VLA-2.0, the embodied agent DR-Agent, the policy optimization framework DR-APO, and the data closed-loop platform DR-Benchmark. The company is also collaborating with seven universities, including MIT, Tsinghua, and Zhejiang University.
Why it matters: The five results span world models, VLA, policy optimization, and data benchmarks, offering a way to compare the division of labor and metrics across the full embodied brain pipeline.
ZhiZhi Innovation Research Institute has published a complete technical report on IQuest-Q1, which has 320B total parameters and 15B activated parameters, disclosing results on 30 benchmarks and training details. The team built more than 1 million executable, containerized task instances from code repositories and existing agent trajectories, and set up a three-layer mechanism of environment preparation, runtime checks, and trajectory auditing to prevent reward hacking.
At the 2026 iFlytek 1024 Developer Festival, Lingdong General released the iFlyBot-WAM industrial world action model and the LindenBot-Harness execution organization system, and debuted the LDU environmental data collection solution. The LDB embodied intelligence general-purpose robot has obtained CR certification and EU CE-MD market access, and in B2B warehouse picking has validated an operational path covering autonomous movement between shelves, recognition of goods in different poses, and flexible grasping. Its data collection, training, and inference robot LDT had previously completed validation in small-item feeding scenarios.
Researchers from Shanghai Jiao Tong University, UC Berkeley, Stanford University and other institutions proposed the Physical RSI framework in the paper "Who Improves the Robot? Toward Physical Recursive Self-Improvement," incorporating task capability updates and improvement process updates into a unified analysis and presenting an L0–L5 grading scale.
The authors propose Sensorimotor Alignment (S²), a training-free offline metric that measures how well a model's sensory representations align with the geometric structure of physical motion. The study finds that S² correlates positively with downstream robot task success across multiple embodied model families, enabling low-cost model screening before closed-loop evaluation and identification of promising VLM backbones before fine-tuning into VLAs.
The paper proposes WarpWeaver, a decentralized real-time planner that negotiates shared-space rents among robots through decentralized column generation, so that the dispersive trees searched by each robot converge to a near-optimal consistent solution. In Dynobench simulations, its total travel time is up to 22% lower than a centralized offline planner and up to 37% lower than decentralized real-time baselines, and on a graph-connectivity task it nearly doubles the success rate of a specialized planner. Onboard planning experiments with six robots verified practical deployability.
The authors propose RGraph-Eval, a reference-demonstration-based framework for evaluating robot videos that uses task scene graphs to explicitly expose interaction errors in generated videos, such as errors involving objects, grasps, and placements. The framework uses a foundation-model pipeline to identify manipulation roles, locate object masks, infer and verify state transitions and action intervals, and reconstruct the 3D motion of manipulated objects. It outputs an interpretable score that covers visible objects, relation changes, and 3D path shape.
LocoManiLab is a unified testbed for humanoid loco-manipulation covering data collection and generation, policy training, evaluation, and real-world deployment. Its task suite includes 234 atomic and compositional tasks across five household scene categories, supported by 51.5 hours of simulated teleoperation data.
The paper proposes the STAMP framework, which uses a "robot-to-human-to-robot" process to connect independently captured humanoid motions into continuous sequences. Its core is a learned alignment model: after motions are retargeted to a human skeleton, the model ranks candidate positions, headings, and transition durations for the subsequent motion, and a human motion prior then generates the transition under the selected alignment. The study evaluates the tracking success rate of the generated transitions on the Unitree G1 and ROBOTIS K1, and demonstrates long-horizon motion transitions in real-robot demos.
The authors propose a spatiotemporal retargeting method, STR, which uses a vision-language model to select relevant motion signals and body parts and identifies keyframes accordingly, allocating additional solver iterations to keyframes and higher interaction weights to the selected parts to reduce retargeting computational overhead.
The study proposes the FRED framework, which predicts future video representations from noise in a single step via self-distillation, using representations extracted from real future observations as targets, while historical observations provide temporal context for the video model. Across diverse real-world manipulation scenarios, FRED outperforms compared world action models, with notable gains on long-horizon and out-of-domain tasks under zero-shot environment generalization; using only human video demonstrations, it achieves a 50.0% success rate on tasks absent from robot training data, versus 19.4% for the best comparable model.
The paper proposes SkeLAM, a skeleton-guided latent action model for controllable robot world modeling. It decomposes latent actions into a kinematic component supervised by the robot skeleton and a residual component that retains complementary visual dynamics, and uses cross-modal swap reconstruction to make RGB and skeleton kinematic representations interchangeable.
At the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference, Unitree Technology founder Wang Xingxing judged that as soon as the end of this year or next year, users may be able to use a single voice command to have a Unitree humanoid robot independently complete household tasks such as tidying a room.
Unitree Technology held the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference in Hangzhou. Founder Wang Xingxing said in a speech that physical AI may see a major breakthrough within one to three years, and said GPT-6 Astra enabled Unitree robots to complete painting and Go-playing experiments under zero-training conditions. He also announced Unitree's in-development Physical AI Robot Self-Evolution V1 system, which lets AI agents write code autonomously, test and train in simulated environments, and validate in real environments, forming a continuous iteration loop; he hopes Unitree robots can enter ordinary households by the end of this year or next year.
A survey examines video foundation models and video world models through the unifying lens of world modeling, organizing the literature into four complementary operations: reading, which infers world properties from observed video; writing, which synthesizes video from specified or learned conditions; sharing, which integrates understanding and generation through common modeling mechanisms; and interacting, which predicts and updates world states under actions and observations.
Using a real-world test report on SenseTime's Astra as its starting point, the article notes that Astra is outstanding at the planning layer (100% on LIBERO-Long++, 81.9% on RoboDojo), but real-machine operation exposes an architectural break between cloud inference's second-level latency and millisecond-level force-feedback closed loops; stacking bowls took an average of about 131 minutes and cost about 19 US dollars.
The IROS 2026 paper NIDAR proposes generating pseudo near-infrared images from RGB images and using intrinsic decomposition to add point-level intensity to existing point clouds; after offline training, it can transfer zero-shot to target scenes. In the Warehouse and Hospital sequences in Isaac Sim, after replacing with NIDAR intensity, the ATE RMSE of ISC-LOAM decreased by about 43.0% and 41.1%, respectively, compared with the simulator's built-in intensity.
Yang Dongsheng, senior vice president of BYD, said at the World Intelligent Connected Vehicles Conference that L3-level conditional autonomous driving is expected to land in 2028, and higher levels may land another 1–2 years after that. Wang Xingxing, founder of Unitree Robotics, said at the 2026 World Physical AI Conference that physical AI is showing signs of being "broken through," with a breakthrough possibly emerging in the next 1–2 or 2–3 years, and that the core challenge of embodied intelligence is aligning AI model inputs and outputs with real robots.
DeepSeek reportedly completed about RMB 100 billion (USD 15 billion) in financing, one of the largest financing deals in China's history, with CATL and Tencent among the largest investors in this round. About 80% of common apps have not completed foldable-screen adaptation for the iPhone Duo; unadapted apps such as WeChat and QQ can only be displayed centered or scaled proportionally, and Tencent customer service said it is actively advancing adaptation and will release updates as soon as possible.
At its Q3 2026 earnings call, Tesla said Optimus has begun production on a high-volume line at its Fremont, California factory, with plans to deliver to external customers around the end of 2027 and to reach an output rate of 20,000 units per week by the end of 2027.
In an exclusive interview with Caixin Time, Xiong Youjun, CEO of the Beijing Humanoid Robot Innovation Center, said the humanoid robot industry has no bubble, only bubble companies, and that the share of prototype robots actually doing real work remains limited. He disclosed that about 80% of Tiangong's data training comes from virtual simulation and 20% from real-world scenarios, that the Pelican-Unify unified representation embodied world model can improve execution accuracy but has high computing power requirements, and that the next funding round is already being planned.
The study proposes a training-free offline metric, Sensorimotor Alignment (S^2), that measures how well a model's sensory representations align with the geometry of physical motion. The authors report that S^2 correlates positively with downstream robot task success across model families, and that it can be used to screen models before closed-loop evaluation and to identify promising VLM backbones before costly VLA fine-tuning.
DiscoDemo uses reinforcement learning as a data generator, training a privileged-information latent-conditioned generator that discovers multiple successful solutions through a parallel reverse curriculum and a success-gated diversity objective, then exports them after safety filtering as a static dataset for downstream visuomotor imitation learning. On four manipulation tasks, the method raises downstream success in simulation from 40% for the strongest baseline to 62%, and on a real robot from 39% to 78%. The authors believe that the diversity of successful solutions can improve imitation learning for a pretrained flow-matching VLA, and that it complements corrective data collection.
The research team proposes Ψ-0.6, a unified vision-language world model that adaptively combines perception, language, geometry, and simulation within a shared reasoning process called streams of thought. These streams alternately generate and interpret images, text, and explicit representations of geometry and motion through a shared token interface. The model surpasses leading open-weight models zero-shot on multi-view spatial reasoning, with further gains from reasoning through geometric intermediate representations.
At the 2026 World Physical AI Conference, Unitree founder Wang Xingxing said physical AI is already showing early signs of being "broken through," with a breakthrough possibly arriving in the next 1-2 or 2-3 years. He believes the core challenge in embodied intelligence reaching its "ChatGPT moment" is aligning AI model inputs and outputs with real robots, and that commercial service scenarios will see deployment sooner than home and industrial ones. In internal tests, he said, GPT-6 deployed on a robot could clean paintbrushes on its own and continue painting without being specifically programmed to do so.
HoloBoxing proposes a simple method for learning humanoid boxing based on a pretrained whole-body controller, separating reusable motor skills from task decisions: a compact planner outputs future body-point targets, while a frozen controller handles joint-level execution.
RynnUnify is a 16B unified embodied model that combines embodied understanding with long-horizon visual planning through interleaved video-language reasoning, covering spatial reasoning, physical prediction, and grounding. The authors also built an agentic extension using reinforcement learning and introduced a world-action extension that translates understanding and prediction capabilities into robot control.
At GTC Berlin, Jensen Huang announced that the Vera Rubin platform has entered full-scale production and introduced DSX POD, a modular AI factory reference design. It consists of 7 liquid-cooled computer racks interconnected by about 17 kilometers of cabling and uses 800V DC power; assuming power and site readiness, deployment time can be shortened from about 6 months to as fast as 14 days, with deliveries expected to begin in early 2027.