AutoTrust AI Lab has open-sourced JEV-27B-VL, a multimodal decision model built on Qwen3.8-27B that extends System 1 decision-making and System 2 reasoning to image inputs, under the Apache 2.0 license.
MiniMax Code has launched M3.1-Flash-Preview, featuring a 1M context window, native support for text, image, and video input, and five adjustable thinking-intensity levels; the company has not published benchmark scores.
We are releasing Dyna-2.1, the first physical agent to achieve reliable super long-horizon whole-body autonomy. It combines our new semi-humanoid hardware with Da...
Being-H0.7, the general-purpose dexterous manipulation large model independently developed by ZhiZaiWuJie, has been accepted to NeurIPS 2026, marking the Being-H series' completion of acceptances to the three top conferences ICML, CoRL, and NeurIPS within one year.
Simplexity Robotics has launched and open-sourced the world-action model SimpleWAM; the 6B-parameter version reduces inference latency on NVIDIA ThorX (T5000) from 181.23ms to 42.31ms, enabling high-frequency closed-loop operation on-device at nearly 24Hz.
Why it matters: The original article presents the on-device optimization path from 181ms to 42ms and four architectural designs, which can be transferred to understand how world-action models can meet high-frequency closed-loop control requirements.
Lingchu Intelligence has officially released its embodied model Psi-R2.5, which uses a two-tier architecture with a QwenVL3.5-4B high-level planner plus a Wan2.2-IT2V-5B low-level controller, and proposes the strong Pair Data standard. The model reverse-invokes the Psi-W0 world model to generate human demonstration videos backward from real-robot data, enabling mass production of paired data, and supports in-context learning (ICL) for zero-shot generalization to new tasks. It comes with a HIL+RL post-training framework and is claimed to raise task success rates to 99% after a few iterations taking only 1-2 working days, and has already been deployed at customer sites.
Why it matters: The original article details the model architecture, the strong Pair Data data standard, and the reverse-generation method, letting readers understand how its data approach supports real-robot generalization.
The 2026 Third Smart Driving "Fengshen Bang" Challenge was held in Changsha. JD Logistics' "Lone Wolf" 6th-Generation Smart Delivery Vehicle Plus Edition won the "Best All-Around Performance Award" with the highest overall score, achieving the best performance in the field across multiple scenarios in its first appearance and passing all difficult scenarios with full marks.
Congratulations to our partner Simate on developing its self-developed Physical RSI system and topping the RoboDojo leaderboard!
Self-play is a powerful path for physical AI. @SkildAI's model trains against increasingly stronger versions of itself, turning a simple objective into emergent behaviors.
Noetix released the technical report for GLOW (Generative Learning of World), a general embodied intelligence generative learning framework, claiming an average success rate of 62.2% across 18 tasks on RoboDojo.
Dexbotic has fully integrated DM0.5, upgrading both its end-to-end pipeline capabilities and its evaluation coverage. Readers can use "Read the original article" to go directly to the GitHub repository and view past issues.
At the Greater Bay Area Film Concert, AGIBOT used the WITA-Omni omni-modal embodied model to achieve real-machine human-robot interaction; the robot can identify its conversation partner, judge the right timing to respond, and coordinate speech, motion, and facial expressions in real time. The model uses a Thinker-Talker-Actor architecture, and WITA-Omni Preview ranked first on the DailyOmni benchmark with an average accuracy of 85.21%, leading in six of eight evaluation dimensions.
Psi Robotics released the Psi-R25 robot manipulation model. Its core upgrades are improved data quality and data attribute compression: it reviewed and extracted high-quality data from 100,000 hours of data, and used strong pair data to align human data with robot data.
Why it matters: The original article details how strong pair data is generated and the data quality standards, giving readers a referenceable path for aligning human data to robot data.
On September 16, the WorldArena 2.0 global final leaderboard was revealed. Koala Uur'an's Youran Boundless World Model ranked second globally in overall score on the Track 1 video quality track, and ranked first in two core metrics: Background Consistency and JEPA Similarity. A total of 80 models were included in this evaluation. The model uses structured 4D world modeling and dynamic scene memory technology, and is being combined with the Geek Mind embodied brain. It has been integrated into a quadruped robot platform and provides an industrial inspection solution for a central state-owned enterprise's manufacturing workshop.
Swinging. Turning. Jumping. @GenrobotAI's DFM converts six-camera ego videos into continuous full-body mesh and skeleton data, maintaining body structure and temporal consistency even during fast motion and complex poses. See the whole body. Learn the whole skill set. #GenRobot #WholeBodyData #egodata #PhysicalAI #Robotics
Figure AI released Helix 2.5, which completed three long-horizon full-body behaviors—tidying a living room, folding towels, and making a bed—zero-shot in 30 Bay Area homes where no data had ever been collected, without collecting data or fine-tuning in the evaluation environments.
Why it matters: The original article provides controlled comparison data on Index pretraining for zero-shot generalization and evaluation details, helping readers understand the actual level of humanoid robot pretraining transfer.
Bend. Squat. Move. DFM can reconstruct continuous full-body meshes from egocentric video, even with partial occlusion. Average full-body error is about 3 cm. See the Whole Body. Learn the Whole Skill. #GenRobot #WholeBodyData #ego data #PhysicalAI #Robotics
The 2026 Pujiang Innovation Forum AI-Enabled Scientific Research Special Forum was held on September 13 at the Shanghai Zhangjiang Science Hall, co-hosted by the Fourth Department of the Ministry of Science and Technology and Shanghai AI Lab. During the forum, Shanghai AI Lab launched the ShuSheng Scientific Multimodal Large Model S2 and the ShuSheng Physical World Model W0, opened the full-stack services of the ShuSheng·DuanYan Scientific Discovery Platform, and jointly released 20 major scientific questions with 14 institutions as well as multiple co-construction achievements of the "Climber Program."
Why it matters: The official report summarizes multiple advances released at the forum, among which the force-touch fusion and robotic arm operation details of the W0 world model may be of reference to readers in the embodied direction.
EBKernel released its brain-inspired cognitive world model Cog-WM 1.0 on September 14 at the Brain-Inspired Embodied Intelligence sub-forum of the Shanghai Pujiang Innovation Forum. It includes a navigation branch, Cog-WM Nav 1.0, and a manipulation branch, Cog-WM Manip 1.0, and uses memory, prediction, and active exploration to help robots understand the physical world.
On September 10, 2026, Suzhou-based JoyIn released the Aether embodied intelligence model, aimed at humanoid robot body control and real-world task execution. JoyIn says the model was trained on about 200 hours of human video and zero real-machine data, has roughly 4 billion parameters, exceeds 90% success rate on first attempts at some tasks, can run on bodies from Unitree Robotics, AgiBot and others, and that it plans to partially open up the model's capabilities.
XPeng XOS 6.3.0 deploys a second-generation VLA upgrade via OTA, restructuring the model architecture around the time dimension: InfiniVLA lets historical visual information continuously participate in current judgments, with the vehicle side selecting 30 seconds of historical context alongside streaming autoregressive inference.
Why it matters: The article explains concretely the trade-offs in co-designing model architecture and deployment under onboard compute constraints, making it a useful engineering reference for deploying large models on limited hardware.
OM-1 has released its first robot foundation model, which can zero-shot generalize to desktop robotic arms, industrial robotic arms, and humanoid robots. The model learns directly from human operation data without using teleoperation or robot data, claims near-human-level dexterity and efficiency, and supports multi-robot collaboration. Chelsea Finn reposted and congratulated the relevant authors on the release.
Junao Panshi released the brain-inspired cognitive world model Cog-WM 1.0 on September 14 at the Brain-Inspired Embodied Intelligence sub-forum of the 2026 Pujiang Innovation Forum, covering the navigation branch Cog-WM Nav 1.0 and the manipulation branch Cog-WM Manip 1.0.
Force Spirit announced that DM0.5 has topped all four RoboColiseum leaderboards (instruction following, spatial understanding, perturbation adaptation, and general manipulation), making it the only model on the leaderboards to rank first in every evaluation.
Why it matters: The original article lists DM0.5's specific scores on multiple leaderboards and its inference latency reduction data, which can be used to assess the practical effectiveness of its pretraining approach.
Tashi Intelligence released AWE3.7, an iterative version of its AWE series general embodied AI model, claiming cross-environment, cross-embodiment, and cross-task generalization, with the same model deployable on dual-arm collaborative robots, wheeled chassis, and humanoid robots. The materials show its performance in industrial scenarios, including parts sorting, tape wrapping, network cable insertion, phone boxing, screw tightening, drive belt grasping, and block sorting, as well as household tasks such as clothes folding, whiteboard wiping, and smoothie making, and claim that model-level changeover time has been compressed to the hour level.
Unitree Robotics announced a full upgrade and open-sourcing of its general-purpose humanoid foundation model, claiming it leads global open-source models on 7 benchmarks. The material provides only a title and a one-sentence summary; it does not give the model name, version number, or specific scores for each benchmark.
Songyan Dynamics has launched Scalabot, an embodied intelligence sub-brand, and released its first world model, HERON-World Model. Building on its strengths in robot hardware, the brand focuses on letting intelligence grow in the real world.
What if humanoid robots could gesture as naturally as they speak? 🤖 Introducing RoboGesture — enabling humanoid robots to listen, respond, and make meaningful gestures. RoboGesture completes the "listen → respond → gesture" loop in about 2 seconds, bringing us closer to natural human-robot interaction. Semantic alignment. Real-time. Safety-aware. 📍 #ECCV2026 · Poster Session 2 · ExHall #178 🗓️ Sep 10 · 16:30–18:30 CEST 🔗 https://robogesture.github.io #Galbot #EmbodiedAI #HumanoidRobot
At the 2026 XAIR Expo Embodied Intelligence Industry Expo, Zhido Future unveiled a technical preview of its differentiable world model "Zhizi World" (PGD World Model) and demonstrated validation progress on its embodied intelligent robot Xiaozhi S3. "Zhizi World" reconstructs particle-based scenes from short multi-view videos and can edit parameters such as object pose, material, mass, and friction to predict the physical outcomes of actions. S3 Xiaozhi features bionic dual arms, lidar-vision fusion navigation, and multimodal voice interaction; on-site demonstrations included scooping popcorn, filling cups and delivering them, and self-service retail shopping.
XPeng held a sharing and experience day themed "TIME," announcing the first major upgrade to its second-generation VLA large model. The new XOS 6.3.0 version will make its global debut on the XPeng G9L, with the core aim of moving AI from 3D spatial understanding to 4D spatiotemporal understanding.
Why it matters: The original breaks the upgrade into three specific changes: contextual memory, reasoning approach, and world model. Readers can use these to understand how the shift from spatial understanding to spatiotemporal understanding is realized.
At the 2026 World Robot Conference, held August 19–23, KEENON Robotics showcased KOM 3.0, a VLA architecture for the service industry, which it says integrates a latent-space world model. On site, the XMAN-R1 autonomously completed the full process from picking up a cup to extracting coffee; two humanoid robots collaborated under unified scheduling to wash and fold clothes; and together with the T10 delivery robot and C40 cleaning robot, they demonstrated "general-purpose + specialized" collaborative service. In the demonstrations, the robots completed long-horizon tasks end-to-end autonomously, with no teleoperation.
In the "Assembly and Loading Station" event of the 2026 World Humanoid Robot Games Industrial Scenario Competition, Lingxi Zhihong's robot completed three tasks fully autonomously: transporting and shelving, picking and loading, and sub-millimeter assembly of 18 engine valves, scoring 84, 65, and 21 points respectively. After a 10-point special deduction, it finished with a total score of 160 points, ranking second only to two teams under a leading industry robotics company in the enterprise rankings, making it the only startup to break into the national top three on its first participation.
X Square Robot has released WALL-SS, a next-scale autoregressive world model that uses a coarse-to-fine, layer-by-layer refinement generation architecture, breaking away from the traditional diffusion-model approach of generating video in a single segment. In testing, its action-following score was 0.29, the only nonzero score among the compared models; the overlap between generated trajectories and action instructions rose from 0.251 to 0.539; streaming inference can reach 60 seconds of footage without collapse; and in closed-loop policy consistency tests across 6 tasks and 5 training stages, the success-rate agreement between virtual and real-world tasks reached 93%.
Why it matters: The source gives three quantitative metrics and the architecture design, which readers can use to assess how feasible the world model is as a policy testbed.
Galbot's G1 robot won gold in all three of WHRG's most challenging scenario competitions: household, food service, and retail. It operated autonomously throughout, with a 100% win rate. The household competition covered tidying, laundry, pick-and-place delivery, and long-horizon tasks that included unexpected interruptions; the food service competition involved understanding voice commands, food handling and heating, drink preparation, and delivery; and the retail competition included picking, restocking, store patrols, and returning misplaced products to their correct positions.
Qianxun Intelligence competed for the first time at the second World Humanoid Robot Games, winning the office-scenario comprehensive service category with 100% model autonomous reasoning and taking second place in the food-and-beverage preparation and sales category.
Qianxun Intelligence announced that its robot Moz1 won the comprehensive service post in the office scenario event at the second World Humanoid Robot Games, becoming the only robot in the field to compete and win using 100% model autonomous reasoning. Moz1, equipped with Spirit v1.6, operated throughout with no teleoperation and zero human intervention, relying on the model to autonomously understand, plan, and execute tasks including venue setup, printing and document delivery, and paper shredding, completing a 20-minute continuous run with a high score.
On August 19, 2026, at the World Robot Conference (WRC), EngineAI unveiled EngineAI Awaken, a human-brain-inspired embodied intelligence engine. It is a five-layer architecture comprising the S1 instinctive reflex layer, the S2 motion planning and generation layer, the S3 cognitive reasoning layer, and the S4+S5 self-growth layer (an emotional large model).
Octopus Dynamics has released SYNWorld V0.5, a general world model for embodied intelligence. It uses a MoT architecture, has a parameter scale of 10+B, and was trained on 300,000+ hours of data covering images, video, and robot action trajectories with action annotations.
Dyna Robotics published how to reproducibly train Dyna-2 on more than 1,000,000 hours of egocentric video.
Skild AI released the robot foundation model S1, which uses video demonstrations as task prompts and can execute manipulation tasks not seen during training without fine-tuning, with long-horizon tasks lasting up to about 10 minutes. At a pretraining data scale of 100,000 hours, the ICL model achieved a 66% success rate on unseen tasks, compared with 9% for the language-prompted VLA; a single demonstration was equivalent to about 380 post-training demonstrations, and the planting task took 11 minutes from recording the demonstration to autonomous execution. S1 was trained on NVIDIA AI infrastructure and has already been put into use with commercial partners.
Why it matters: The original article provides comparison figures for ICL and VLA at the same data scale and a deployment timeline, allowing readers to assess the feasibility of the in-context learning approach for manipulation tasks.