NaiveAI has open-sourced Naive-N0.5-Flash. Users only need to write a one-page program.md task brief, and an AI agent connected to Claude Code can then complete the entire workflow of data cleaning, training, evaluation, and troubleshooting.
Tsinghua University and other institutions have proposed the ART (Agentic Robot with Tool-use) framework, which can fine-tune any end-to-end VLA model into an agent capable of plug-and-play use of external tools. The method uses 30,000 tool-call trajectories and two-stage LoRA fine-tuning, achieving average success rates of about 75% in LIBERO simulation and about 62% on a real Astribot S1 robot, an improvement of about 20% over baselines such as OpenVLA and π0.
Amodal3R at ICCV 2025 performs occlusion reasoning and geometry generation directly in 3D latent space, using visible-region masks and occluded-region masks to preserve real observations and infer occluded geometry and appearance.
WorldFoundry proposes open-source infrastructure for world intelligence, intended for reproducible inference and evaluation across video, 3D and 4D, embodied action, and interactive world systems. The work defines a generative world model as a predictive dynamical system that maintains a persistent, history-conditioned state representation, responds to actions or interventions, and generates temporally coherent trajectories. Its catalog includes 270 model entries, 74 benchmarks, and 62 metric entries, and it records execution readiness and evaluation conclusions separately through manifests, artifact contracts, and evidence-aware scorecards.
At the World Internet Conference in Wuzhen, Moxin Technology released MoWorld 2.0, a new-generation general-purpose 4D world model that integrates more than ten physical elements, including gravity, elasticity, buoyancy, lighting, soft bodies, fluids, rigid bodies, deformable bodies, and thermal dynamics, into a single model to more accurately simulate object motion, collisions, and environmental changes.
Reinforcement learning policy optimization has evolved from REINFORCE and Actor-Critic to TRPO and then PPO, with the core goal of making policy gradient updates more stable and efficient. PPO uses clipping on the probability ratio between the new and old policies to limit the size of each update, balancing implementation complexity and stability, and is suitable for both continuous and discrete actions. The article also reviews GAE advantage estimation, the on-policy data flow, and tuning approaches for key hyperparameters such as learning rate and clip range.
Nuoyin Intelligence says KnowinAgent achieved a team self-tested composite score of 87.60 and an overall success rate of 82.83% across all 54 task variants in RoboDojo, once again ranking TOP1 on the leaderboard.
Lingji Robotics announced an upgrade to its embodied intelligence model service platform Dexmal MaaS, with a focus on opening post-training capabilities for real-world scenarios. Developers can upload robot manipulation data and complete training and inference deployment in the cloud based on DM0.5.
Apple released the HALO humanoid whole-body motion control system, which procedurally generates 1,000 virtual residential scenes and synthesizes nearly 300 hours of paired data, enabling the Unitree G1 to master skills such as sitting on chairs, turning sideways to pass through a half-open door, crouching under a table, and pushing boxes.
Why it matters: Apple's HALO uses simulation-synthesized data to train humanoid whole-body skills, transfers them zero-shot to real scenes, and ablation experiments show the contribution magnitude of each module.
NVIDIA has released VFlash, a general-purpose decoding framework for fast visual planning that replaces token-by-token coordinate generation with parallel decoding of complete visual primitives, allowing it to predict robot grasp positions and motion trajectories in one pass. At batch size 1 on H100, single-generation speed improved by about 25x; across four MolmoSpaces tasks, the world action model averaged a 70.2% success rate versus a 62.6% baseline. Jim Fan shared that it achieves up to 12x faster action decoding, 5.1x overall robot inference speedup, and up to 61% lower 3D trajectory error.
Why it matters: NVIDIA's new decoding framework changes visual planning from token-by-token generation to parallel prediction, and provides speedup and success-rate comparisons across multiple robot and autonomous-driving benchmarks.
Daxiao proposes a "dual-brain collaboration" architecture that connects the cloud brain SenseNova's task understanding and planning capabilities with the on-device Kaiwu world model's judgment of physical states and prediction of action consequences, while an embodied Harness organizes perception, memory, skills, and control. The exploration aims to preserve a single robot's interaction experience, along with evidence and applicable conditions, so that other robots can reuse it across tasks, scenarios, and embodiments, reducing repeated trial and error. The system is organized around three progressive loops: autonomous execution, experience learning, and recursive self-improvement. It is currently at the exploration stage and has not yet provided real-machine validation metrics.
On November 5, the 'Triple T' Zhangjiang Summit technical salon, co-hosted by an ecosystem company of Zhangjiang AI Innovation Town and Synced, was held under the theme 'The Next Data War in Embodied Intelligence,' featuring a two-person technical debate, three Open Mic sessions, and one industry panel. Huang Yongtao of Ant Lingbo Technology argued that at this stage priority should be given to accumulating real-robot data and incorporating touch as early as possible, while Liu Miao of Tsinghua University held that data selection depends on application positioning and the training stage. Ma Daolin of Qianjue Robotics demonstrated the performance of a VTLA model integrating vision, touch, language, and action in real-robot tasks such as automatic shoelace threading and carton forming.
Yuance Future has released Anchor 1, a whole-body control model positioned as the System 0 cerebellum layer within a whole-body intelligence architecture. It uses five anchor points—the pelvis, left and right wrists, and left and right ankles—as commands and does not rely on online retargeting.
Why it matters: The original article provides quantitative comparisons of the whole-body control model's real-machine error and teleoperation collection efficiency, which can be used to assess how the cerebellum layer's precision affects the success rate of upper-layer VLA tasks.
A team from Northwestern, CMU, UNC Chapel Hill and Amazon, in the paper "WOVEN: Weaving Visual World Modeling into Multimodal LLMs," defines world modeling as estimating the state transition p(s′|s,a) and builds WOVEN, a training and evaluation base of 36K questions.
Why it matters: The paper breaks world modeling into independently intervenable state-transition tasks and provides reusable rules for selecting training data, offering reference value for multimodal model training.
The Hong Kong University of Science and Technology, Beijing Academy of Artificial Intelligence, Peking University and other institutions jointly proposed DexRoam, which uses first-person human whole-body demonstrations to train mobile dual-arm dexterous manipulation policies. The work has been accepted to CoRL 2026. The system narrows the human-robot motion distribution gap through three stages: embodiment alignment, action semantic alignment, and temporal alignment, reducing MMD and SWD metrics by an average of 41.7% and 31.1%, respectively.
World Labs, founded by Fei-Fei Li, has launched its first app, Frameboy, which uses the Atlas world model to generate first-person scenes in real time, streamed via cloud GPUs at up to 15 fps, with measured response of about 70ms.
miHoYo, together with Tsinghua University, the University of Hong Kong, the National University of Singapore, and other institutions, has open-sourced the GameScaling series of general game models, covering four size tiers—0.8B, 2B, 9B, and 27B—for a total of eight models. The weights and inference code are released under the Apache 2.0 license.
Dream Labs, spun out of NVIDIA's GR00T team, has completed a $90 million seed round and launched Dream API, a publicly registrable, explicitly priced robot foundation model API service.
Why it matters: Dream Labs prices robot brains by the hour, giving readers a clear view of how embodied intelligence is shifting from project-based pricing to an on-demand cost structure.
AgiBot launched a 24-hour convenience store work challenge in which 6 Genie G2 robots took on roles including reception, pickup and delivery, drink preparation, inspection and restocking, and cleaning, completing 1,730 tasks in total; in the end, 4 passed probation and 2 did not. The robots run on the embodied Agent framework Genie Master, which uses Physical RSI to turn success and failure experience from execution into long-term memory and iteratively refine action strategies; capabilities such as judgment for restocking and cleaning come from this prior self-review.
NEURA Robotics has entered a partnership with E.ON and the enviaM Group to jointly develop Physical AI for critical energy infrastructure, for inspection, asset monitoring, and assistance with high-risk, physically demanding, or repetitive work.
Shanghai Kepler has built a closed loop from simulation training to real-machine validation for industrial humanoid robots based on NVIDIA Isaac Lab, Isaac GR00T, and Cosmos. The company develops its own planetary roller screws, linear and rotary actuators, joint modules, and robot bodies, introduces real dynamics parameters into Isaac Lab for Sim2Real iteration, and validates them in scenarios such as workshop assembly and CNC machine loading and unloading. Kepler is also working with a team from Zhejiang University's College of Artificial Intelligence to develop an embodied intelligence foundation model for long-horizon tasks.
On November 3, Zhijian Dynamics and Hangcha Group completed the deployment of the first robot sorting line for a supermarket front warehouse, connecting the full workflow from AMR merge feeding, robot autonomous planning and dexterous grasping and turning, and scan verification to order distribution. The line uses the iX robot body that Zhijian Dynamics customized for Hangcha, and relies on visual perception and VLA task planning to handle boxed, bagged, bottled, and other product forms without the need to model each SKU separately. Measured single-machine sorting efficiency reaches up to 450 items per hour, with accuracy stable above 98%, and under full load it can support 24-hour continuous fault-free operation.
Why it matters: Readers can learn about the actual deployment conditions of the supermarket front-warehouse sorting line and the single-machine efficiency and accuracy data.
The Beijing Academy of Artificial Intelligence (BAAI) proposes ArtHOI, which uses video priors to achieve high-precision reconstruction of 4D articulated human-object interaction without paired 3D data. The method targets the human-object interaction reconstruction task, and its core idea is to replace paired 3D data with video priors.
Geek+ announced a 24/7 warehouse autonomous operations plan, aiming to keep warehouses running around the clock and continuously optimize through real operational experience. The plan relies on its end-to-end robotics product portfolio, the Geek+ Warehouse Brain software platform, and its global deployment and service network, and uses the Gravity unified embodied AI framework and the Gravity 4D embodied world model to extend automation from structured handling to more variable operations and exception handling.
Six AGIBOT G2 wheeled humanoid robots completed a 24-hour trial at an operating convenience store in Xuhui District, Shanghai, performing a total of 1,730 tasks, with four of them passing the final assessment.
Why it matters: AGIBOT disclosed the task distribution and assessment results of six G2 robots over 24 hours in a real convenience store, offering a view of how embodied intelligence performs in continuous operation.
In Q3 2026, Unitree's revenue was 665 million yuan, up 69.93% year on year, while non-recurring net profit was 105 million yuan, down 17.42% year on year. R&D spending was 127 million yuan, up 249.58% year on year, showing revenue growth without profit growth.
Shanghai AI Lab has launched its laboratory operating system Sage·LabOS, which uses Sage SCP-Lab as a unified interface and the Sage physical world model W0 to drive robots in completing experimental operations, connecting heterogeneous devices for autonomous execution and data return. The system has four core capabilities: capability discovery, task orchestration, execution scheduling, and data feedback. In a lipid nanoparticle (LNP) formulation optimization task at Lingang Laboratory, it completed five rounds of dry-wet closed-loop iterations, automatically synthesizing and validating 335 formulations, with encapsulation efficiency rising from 67% to 98%.
Why it matters: The source provides the four core capabilities of the autonomous laboratory operating system and measured dry-wet closed-loop data from LNP formulation optimization, showing how robots can be integrated into scientific research workflows.
The study proposes SLA, a spatial-language-action model initialized from a pretrained pi05 policy, which jointly optimizes action prediction and the alignment of intermediate visual embeddings with the spatial representations of a frozen 3D perception model during continual pre-training. The team unified heterogeneous robot datasets spanning embodiments, tasks, and scenes, and introduced feature-dependent gates that adaptively aggregate visual embeddings from multiple layers; the teacher model can use partially available depth and camera calibration information. All auxiliary components are used only during training, and the inference architecture remains unchanged.
German retail group Otto will deploy AI dual-arm robots in phases at a logistics center in Poland to inspect, identify, and sort e-commerce returns. The first phase plans to deploy 10 systems, and whether 15 more will be added later depends on initial operating results. Returned goods are often crumpled, repackaged, or presented in various orientations, so the robots need to combine visual information to decide how to grasp and adjust their actions based on feedback; simulation can model different orientations and packaging to assist training. Currently, the project still lacks enough public data to judge actual efficiency and return on investment.
The article traces the architectural evolution of three generations of robot foundation models: RT-1, OpenVLA, and π0. It focuses on the architectural changes across these three generations and is a technical evolution analysis rather than the release of a new model.
Shanghai EABOT Technology (EABOT) recently completed two consecutive funding rounds totaling nearly 400 million yuan, with investors including Ge Weidong, Chentao Capital, Hony Capital, and Zero One Automotive. Founded in February 2026, the company focuses on heavy-duty industrial physical AI. It already operates more than 100 heavy trucks across 60 routes, mainly serving short-haul bulk material scenarios of 0-300 kilometers, and has conducted real-road zero-intervention testing in Xinjiang, Inner Mongolia, and other regions.
Beijing Humanoid released the Hi-WBC general whole-body control model on the "Huisikaiwu" platform on November 2. The model combines human motion references with real physical constraints, enabling robots to adjust how they execute tasks as payload, external forces, and terrain change. It uses a continual skill expansion framework based on MoE and policy distillation, Contact-aware Learning that incorporates contact force prediction, and Perceptive Control that introduces terrain awareness.
Why it matters: The original text discloses Hi-WBC's technical approach to handling payload, external force, and terrain constraints beyond motion tracking, offering insight into how whole-body control is moving from demonstrations toward real work.
The research team proposes A2-MEAS, an embodied multi-agent framework that uses code as a scaling interface. It combines correction codes with corrective language, spatial evidence, execution history, and action proposals, while MetaBrain provides training supervision and runtime guidance.
Why it matters: The framework uses code as a scaling interface, injects corrective language and execution history into the policy, and presents comparison data across multiple benchmarks and real-robot experiments.
AgiBot launched a convenience store livestream challenge on November 2, with 6 Genie G2 robots on duty continuously for 24 hours. When it ended on November 3, 4 had successfully passed probation and 2 had not yet passed. During the assessment period, the robots completed 1,730 tasks/items in total, including receiving 683 customers, picking up and delivering goods for 369 orders, making 147 cups of coffee, producing 126 servings of hot food, carrying out 277 inspection and restocking rounds, and clearing tables 128 times.
At CeMAT, Geek+ exhibited the general-purpose humanoid robot Gino 1, the RoboShuttle Hyper case-to-person system, and the SkyCube pallet-to-person system, and demonstrated a comparison of prediction and execution by its self-developed Gravity embodied model in clearing bins of mixed items and picking unfamiliar items.
On November 5, Tesla showcased its first production version of the humanoid robot Optimus at the 9th China International Import Expo, presenting its deployment directions through scenarios such as carrying luggage, dog walking, and factory sorting. Tesla said the mass-production line at its Fremont factory has started production, and the units rolling off the line are commercially viable versions, but at this stage they will be kept for internal factory use for material handling, sorting, and inspection, and will not be sold externally for now. At the Q3 earnings call, Musk said it will take at least more than a year to reach a steady-state capacity of 20,000 units per week, that the long-term target for per-unit cost is $20,000 to $30,000, and that the biggest current engineering bottleneck is the dexterous hand.
LiberAI, founded by Liu Songming, has released Liber-0, a new-generation world action model that uses three expert streams for language, vision, and action along with Multimodal Attention to jointly predict future frames and actions.
Why it matters: The technical report uses controlled experiments to break down the contributions of human data, action supervision, and embodiment diversity to dexterous manipulation, which can inform judgments about data-mix strategies.
Sun Yat-sen University, Peng Cheng Laboratory, and other institutions propose RePO-VLA, which segments recovery trajectories, trains a progress-aware semantic value function, and performs value-conditioned fine-tuning so that dual-arm VLA policies can continue advancing tasks after disturbances. The study also builds FRBench, a failure-recovery benchmark; the RoboTwin simulation contains 46 dual-arm tasks and 23,453 trajectories. On real robots in water-pouring and towel-folding tasks, as recovery demonstrations per task increase from 50 to 200, the average disturbed success rates are 30%, 65%, and 75%, respectively.
A team from Xiamen University and Nanyang Technological University proposed RoTA-Bench, which uses interactive 3D-to-2D projection alignment to evaluate whether vision-language models can precisely adjust an object's 3D pose based on visual feedback. The benchmark contains 674 tasks, of which 100 form a frozen test set; models must choose among 24 discrete rotation actions so that the projection reaches an IoU of 0.90 with the target contour.
The article argues that, limited by onboard compute and heat dissipation in robot bodies, embodied model inference is shifting to the cloud, and embodied AI MaaS is becoming a new direction. In July 2026, Yuanli Lingji released a general embodied MaaS service; in October, Physical Intelligence launched an invitation-only Pi API Beta; and Dream Labs launched Dream API backed by a $90 million seed round.