TypeSafeAI released the decision model Jev, which outputs only yes/no, option numbers, or probability scores. Within weeks of launch, users exceeded one million, about one-third of Fortune 500 companies began using it, and its valuation rose from about $200 million to $7.5 billion in 24 days.
Addressing the problem of being unable to train a policy from HDF5 demonstration data, the article outlines 8 critical details. It focuses on quality and format issues in the demonstration data itself and explains how these details affect policy training outcomes.
Flex-π targets dual-arm robot manipulation, using a single approximately 6-billion-parameter model to jointly learn actions, future appearance, 3D geometry, and object semantics, then selecting at deployment which content to generate. It addresses the tension that predicting the future can provide more information, yet not all tasks require all prediction content.
BSC-Nav is an arXiv paper (2610.02784) aimed at indoor object navigation, long-instruction following, active embodied question answering, and mobile manipulation, addressing the problem that multimodal large models can understand the current view but struggle to support spatial cognition. This analysis article from Guyuehome says it moves embodied intelligence from reacting to a single glance toward retrievable spatial cognition, but currently only the abstract is provided; method details and experimental results are not elaborated.
The article reviews three paths for embedding physical priors into robot learning: data and representation, model architecture, and training loss, and emphasizes that where a prior is placed determines its constraining effect. Using grasping a soft object as an example, it explains the layered roles of kinematics, dynamics, contact, and material properties in the task, while noting that physical priors are not a panacea and need to match the task.
SSTG-Nav proposes a reusable object navigation framework for long-term deployment scenarios, focused on addressing the mismatch between observation poses and docking poses. The framework uses a single task-agnostic RGB-D mapping pass, employing a vision-language model (VLM) for detection and depth back-projection to generate reachable targets. The material provides only an abstract and does not give full experiments or results.
The author loads an open-source pretrained policy in Isaac Sim to reproduce quadrotor racing on a figure-8 track, matching the official result with 25 gates and 3.12 laps. The content breaks down the three attitude phases of high-speed turns and uses a three-seed controlled experiment to test the real benefits of domain randomization.
Future Not Far released the Self-Evolving WAM self-evolving world-action model, integrating joint world-action modeling, interactive experience learning, and training-effect feedback into a single system. In update attempts from round 0 to round 30, all four groups of household tasks showed gains in autonomous full-completion success rates: folding clothes rose from 54.6% to 96.8%, hanging clothes from 46.2% to 91.2%, making pasta from 42.7% to 89.5%, and stir-fried pork with chili peppers from 38.2% to 68.9%.
Why it matters: The original text gives the success-rate changes for four groups of household tasks from round 0 to round 30 updates, showing the practical benefits of the self-evolving training loop.
The Lion Rock AI Lab under the China Merchants Group Advanced Technology Research Institute has released an embodied reasoning vision-language model, Lion-ER-35B-A3B. It uses a MoE architecture with about 35B total parameters and about 3B activated per token. On ten public embodied-related benchmarks, all metrics improved over the base model, with an equal-weight composite average score of 65.29.
World Labs launched its next-generation world model Atlas on September 1. It is currently not open source and is available for early trials only to a small number of partners. Atlas takes 1 to 6 reference images and a camera trajectory as input, can generate videos up to 1 minute long at 1440p, and exports point clouds and 3D Gaussian splatting, used to reconstruct real scenes as robot simulation training environments. In the official demo, 24 frames taken from a phone video were used to reconstruct a warehouse-sized space, allowing a virtual robot to walk through it and generating camera images and depth data; however, there is no public data yet on complex physical behavior or large-scale operating costs.
At CeMAT ASIA 2026, Seer Robotics exhibited its Robot Brain product line and several robots equipped with the platform, including the X1 PRO humanoid robot autonomously completing a water-fetching task, the P300 and P1500 AI delivery robots collaboratively avoiding obstacles, and an omnidirectional stacker forklift autonomously identifying and picking up various pallets. Robot Brain consists of three parts: Model, Cerebellum, and Neural pathways. It has been adapted to more than 400 types of core components and supports more than 2000 robot SKUs.
Sodero's co-founder demonstrated a humanoid robot walking while avoiding obstacles, and discussed how to test trained policies more realistically before deploying them on real hardware. The team previously deployed a dancing policy on the Unitree G1; it worked reasonably well, but the robot sometimes fell. The team is exploring system simulation so that the simulation reproduces how the policy receives sensor data, how it issues commands, and physical characteristics, in order to test policies under conditions closer to real deployment.
LiberAI has released Liber-0, a world action model pretrained on more than 2.2 million hours of human experience, combining large-scale human video, hand motions recorded with data gloves, and robot demonstrations, with the goal of giving robots generalizable manipulation capabilities. In an eight-task evaluation, a policy trained with 25% robot demonstrations plus 100% glove demonstrations scored 75.6 on progress, higher than the 72.8 scored using only all robot demonstrations. In a 60-task zero-shot evaluation, Liber-0 achieved a 36% success rate.
The study proposes Route Anything, a unified framework for studying model and policy routing in general-purpose AI systems, and establishes an optimal routing theory under quality-cost trade-offs. The authors construct an omni-modal trajectory matrix covering 25 benchmarks across language, image, video, speech, coding and workplace agents, and robotic manipulation, with 452 candidate configurations and 374,401 scoring runs.
The research team proposes PrimeSim, an action-conditioned multiview world model simulator, midtrained on robot demonstration data covering 22 platforms and 30,110 hours, and has released the model weights, inference code, and training code.
Why it matters: This work presents a world model simulator trained on demonstrations spanning 22 platforms and 30,000 hours, and releases the weights and training code, facilitating comparison of its long-horizon stability and action-following capability.
Warehousing and logistics is seen as the scenario closest to real production for embodied intelligence, because it combines high-frequency real demand, a semi-structured environment, and quantifiable business metrics. The article divides the players into three groups: embodied intelligence startups, traditional logistics equipment companies, and end logistics enterprises such as JD Logistics, SF Express, and China Post, which respectively hold technology, systems, and scenario resources.
DreamFree proposes a world action model (WAM) that does not require generating future videos. Through representation distillation, it turns a pretrained video diffusion model into a predictive visual encoder that predicts the frozen video model's representation using only the current observation; an action module then extracts action-relevant information.
Tsinghua University and UC Berkeley jointly proposed and open-sourced uni-robot-agent, a cross-environment unified embodied agent harness for connecting tool interfaces across different robots and simulation environments. Based on this framework, Qwen3.8-Flash-Next scored 31.39 on the RoboDojo benchmark, exceeding the publicly reported 28.97 for GPT-6 Astra, at about 1/20 of Astra's cost.
Liu Shilong, a postdoctoral researcher at Princeton (joining Columbia University in fall 2027 as a tenure-track assistant professor), has shifted his research focus to self-evolving agents, with robotics and scientific discovery as application scenarios. He is a core author of the open-world detection model Grounding DINO, which has over 2 million monthly downloads on Hugging Face, and whose open-source project has accumulated over 30,000 GitHub stars.
On November 2, six AgiBot Genie G2 humanoid robots worked 24 consecutive hours at a normally operating convenience store in Xuhui, Shanghai, covering roles including greeting customers, making coffee, stocking shelves, cashiering and cleaning, with the entire shift livestreamed. The six robots shared task scheduling through AgiBot's embodied Agent framework Genie Master and reported execution results back via the embodied execution protocol EEP, collecting failure cases into an error log and converting them into training data.
At CeMAT Asia, which opened on November 3, Ingsmart exhibited the Messol embodied-intelligence loading and unloading robot solution, designed for native mixed-box conditions inside truck trailers; it supports grasping of mixed boxes in multiple box types and orientations, and the operation process requires no manual entry into the trailer for sorting.
At CeMAT 2026, Geek+ demonstrated a robotic arm equipped with the Gravity embodied model and its general-purpose humanoid robot Gino 1, completing grasping of unfamiliar objects, picking, and bin handling. Gravity uses a hybrid Transformer (MoT) dual-brain collaborative architecture and a hierarchical memory mechanism, and through natural-language correction it raised the overall success rate of complex manipulation tasks from 17.5% to 87.5%, with a single picking action stable at about 2 seconds.
NTU MMLab and Shanghai AI Lab propose PhysLDM, which uses a spatiotemporal VAE to compress an entire deformable-object simulation trajectory into latent space and then generates it in one pass with a diffusion model, achieving an average reconstruction error of 2.48 millimeters.
Starting at 10:00 on November 2, six AgiBot Genie G2 humanoid robots worked 24 consecutive hours at a normally operating convenience store in Xuhui, Shanghai. They were assessed against store standards including on-duty time, task success rate, and response time. In the end, 4 passed probation, while Lengjie and Juanwang did not yet pass probation due to shortcomings in on-duty initiative and job stability.
The paper LT-Mem from the APRL lab at Daegu Gyeongbuk Institute of Science and Technology (DGIST) in South Korea won the IROS 2026 Best Paper award. The work proposes a volatile-perception spatiotemporal memory framework that uses three memory layers—Live, Delta, and Meta—and three volatility-based update strategies—coverage, retention, and multi-hypothesis—to solve the temporal forgetting problem in which new observations overwrite old memories when a robot revisits the same space multiple times.
Starting November 2, six Genie G2 robots equipped with AgiBot's Genie Master embodied Agent framework worked continuous 24-hour shifts in a real convenience store. In the end, four passed the probation assessment, completing a cumulative 1,730 tasks, including greeting customers 683 times, fetching and delivering 369 items, and making coffee 147 times.
An ETH Zurich team has developed FlexiQuad, a ring-shaped flexible-frame quadrotor with no central body; on impact the frame deforms to absorb energy and the drone can keep flying. It reaches a top flight speed of 80 km/h, with collision tests at about 25 km/h; the frame material costs less than 15 Swiss francs, and it has been used to grasp and carry soft first-aid kits and payloads heavier than its own weight. The team also used reinforcement learning to let it stabilize its attitude after a collision before resuming autonomous flight, but reliable perception and autonomous navigation remain prerequisites for leaving the lab.
Shanghai Sage Intelligent Technology Co., Ltd. has completed a pre-Series B funding round of hundreds of millions of yuan. Investors include Jiantou Investment, Nantong Investment Management, and Yunqi Capital. The funds will be used for large-model R&D, iteration of the multi-arm robot body, production line expansion, and global market expansion.
Zhicheng AI announced the completion of a funding round of hundreds of millions of yuan, jointly invested by Hangzhou Capital, Fuzhe Technology, Caitong Capital, and Linglan Capital; the funds will be used for embodied intelligence and world model R&D and commercialization. The company was founded in March 2024 by Hu Luhui, former chief engineering lead at Meta. Its Chengling physical world model adopts the JEPA architecture, and version 0.1 was open-sourced under the MIT license in May 2026. The XR1 composite wheeled industrial robot has entered real production line operations at NIO's factory, while cooperation with Volkswagen remains in the scenario testing stage at the final assembly section.
Zou Danping's team at Shanghai Jiao Tong University proposes DiffDreamer, which couples a learned world model with differentiable simulation for agile quadrotor control; the paper has been accepted to RAL 2026. The method uses the world model to predict external dynamics and differentiable simulation to propagate the robot's physical state. In aggressive tracking tasks, distance error drops to 0.381 m, versus 0.793 m for the strongest baseline; in dynamic gate crossing at gate speeds of 1.0/1.5/2.0 m/s, the number of gates passed is 16/12/8, versus 12/4/1 for MPC.
This article reproduces three official pretrained policies for the Revo3 dexterous hand in Isaac Lab for in-hand object reorientation control. The cube random-target reorientation runs 8 times per episode, while the cylinder uses a 22.5° incremental chain, advancing once every 3.4 steps.
PHYMI released its first Physical Agent action partner, Sagi, on October 28. It uses a four-wheeled legged configuration with a humanoid upper body and a companion cargo box, and is positioned to complete long-range mobility and object manipulation in open, dynamic real-world scenarios.
The Xbotics Embodied Intelligence Lab built a closed-loop simulation pipeline using the AgiBot GE-Sim 2.0 world model and the OpenPI π0.5 policy. In testing, the world model generated about 4.33 frames per second, the pure-policy closed loop ran at about 3.20 frames per second, and after adding MiMo online scoring, the rate dropped to about 0.88 frames per second.
Why it matters: The author uses a four-stage test to break down the frame-rate bottlenecks and causes of reward hacking in world-model closed-loop simulation, which can be transferred to the design of VLA evaluation workflows.
Spirit AI's Spirit v1.6 embodied foundation model won the "New Shining Scenario Award" at the 2026 World Internet Conference "Light of the Internet" Expo. The award recognizes digital-intelligence benchmarks with solid results and scale-up potential; Spirit AI has already deployed in real production lines, with partners including Bosch, JD.com, CATL, and Schaeffler.
XPENG ROBOTICS proposes the embodied vision-language model Capek 0.5, training separate experts for four execution capability categories—spatial reasoning, temporal understanding, action guidance, and state verification—then integrating them into a single model via TIES merging and routed MOPD distillation.
Lingchu Intelligence and Hengyan Technology are jointly delivering the Li-Ning East China Smart Logistics Center project. The embodied robot, equipped with Lingchu's self-developed model, has been deployed on site and entered trial operation, taking on routine work and fulfilling commercial orders. The solution targets the bin-picking scenario in footwear and apparel warehousing, where shoeboxes are mixed in storage and item poses are random. The model can autonomously break down multi-step goals and complete recognition, pose estimation, dexterous grasping, product sorting, and pushing full boxes, without requiring large-scale modification of existing infrastructure. During the trial operation, Lingchu will work with Li-Ning's warehouse operations team on stress testing and validation of abnormal conditions, optimizing bin-picking success rates and stability during long periods of continuous operation.
Lingchu Intelligence, together with Beihang University, Peking University, and The Chinese University of Hong Kong, proposed a frequency-aware flow matching method called FAFM; the paper was accepted to the NeurIPS 2026 main conference. The method uses the discrete cosine transform (DCT-II) to represent actions as frequency-domain coefficients corresponding to real time, performs flow matching in coefficient space and constrains the first-order time derivative, requires no additional learnable parameters, and can be used independently or integrated into VLA models such as π₀ and π₀.₅.
Why it matters: The paper presents a frequency-domain action representation that adds no learnable parameters and can be integrated into existing VLA models, offering a transferable approach for joint training on heterogeneous-frequency data.
RoboFFT, proposed by the PKU-Spirit AI Joint Lab, has been accepted to CoRL2026. It is a reinforcement learning post-training method for generative robotic policies that reuses the forward noising process from diffusion model and flow model pretraining to construct an approximate probability ratio for PPO-style clipped updates.
Why it matters: The method reuses the forward noising loss of diffusion and flow models to construct policy update signals, offering a reusable approach for post-training generative policies.
Xiaomi Group received the Outstanding Contribution Award at the 2026 World Internet Conference, and its "key technologies for a general-purpose robot manipulation foundation model for large-scale deployment" received the Leading Science and Technology Award. Xiaomi also disclosed that its humanoid robot took on an 8-hour shift for the first time at a center-console side-cover sequencing workstation in a car factory and achieved autonomous anomaly correction; an embodied foundation model based on Xiaomi-Robotics-1 began trial operation at Xiaomi Youth Apartments, completing long-horizon task validation of over 30 consecutive minutes, covering 3 rooms and 50+ tasks.
Zhizhi Pingfang released NeuroVLA, a brain-inspired embodied intelligence system designed with a three-layer structure of cortex, cerebellum, and spinal cord; the spinal cord layer is driven by a spiking neural network, consumes 0.4 watts, and completes a reflex within 20 milliseconds after a collision. Qianxun Intelligence's Spirit v1.6 ranked first overall in the RoboArena evaluation, surpassing NVIDIA Cosmos3 and Physical Intelligence's π0.5. Leading companies have formed four routes—world-model-native, VLA and world model fusion, brain-inspired architecture, and unified model—but there is still no unified classification standard.