Infinigence AI with Tsinghua and AgiBot proposes collective experience learning CEL, with a 98.3% success rate validated on 96 robots
This month world models became the main line of model releases, with Google DeepMind launching the real-time interactive Genie 4, and Qiyuan, Ropedia and others also successively releasing world models for robot manipulation and cross-embodiment applications. Embodied models and development tools were updated intensively: Zhiyuan released GO-3 Preview and WITA-OMNI, and Pokebot Robotics launched an out-of-the-box embodied API. At the industry level, Unitree Robotics' market value retreated sharply after its listing on the STAR Market, and Wang Xingxing reflected at the Industry Expo on insufficient investment in AI. In research, Collective Experience Learning CEL was validated on 96 robots, with the success rate rising to 98.3%.
This month world models became the main line of model releases, with Google DeepMind launching the real-time interactive Genie 4, and Qiyuan, Ropedia and others also successively releasing world models for robot manipulation and cross-embodiment applications. Embodied models and development tools were updated intensively: Zhiyuan released GO-3 Preview and WITA-OMNI, and Pokebot Robotics launched an out-of-the-box embodied API. At the industry level, Unitree Robotics' market value retreated sharply after its listing on the STAR Market, and Wang Xingxing reflected at the Industry Expo on insufficient investment in AI. In research, Collective Experience Learning CEL was validated on 96 robots, with the success rate rising to 98.3%.
Highlights
Google DeepMind releases world model Genie 4, supporting real-time interaction and native audio
Google DeepMind released its new-generation general world model Genie 4 on the evening of the 28th, which can generate and modify interactive scenes in real time based on prompts, maintaining a single coherent world for about 2 minutes at about 24 frames per second.
36氪 · 微信公众号Source ↗Harvard, NVIDIA and others propose RAX: a robot agent explores online, leading by an average of 37.4 points across five major benchmarks
Research teams from Harvard, NVIDIA, Berkeley, MIT and Stanford proposed RAX (Robot Agent eXploration), which connects a general coding agent to robot perception and control interfaces, and through online exploration writes successful decisions into executable programs and accumulates them as skills, while model parameters remain frozen.
CVer · 微信公众号Source ↗WorldOracle releases an action-conditioned world model benchmark; seven models show deviations between predictions and real execution
Princeton University, the University of Texas at Austin, the University of Toronto, NVIDIA and AgiBot jointly launched the action-conditioned world model benchmark WorldOracle, using the same initial frame and action sequence to compare model predictions with real execution.
The Robot ReportSource ↗
Model Releases/Updates
World models and embodied models made a dense showing, with multiple new products emphasizing real-time interaction, cross-embodiment generalization and full-pipeline scalable learning, while open source and open interfaces became a common choice.
Google DeepMind releases world model Genie 4, supporting real-time interaction and native audio
Google DeepMind released its new-generation general world model Genie 4 on the evening of the 28th, which can generate and modify interactive scenes in real time based on prompts, maintaining a single coherent world for about 2 minutes at about 24 frames per second.
36氪 · 微信公众号Source ↗AgiBot releases embodied model GO-3 Preview, connecting full-pipeline scalable learning from pretraining to post-deployment
AgiBot released the embodied model GO-3 Preview, focused on full-pipeline scalable learning across pretraining, post-training and post-deployment, and ranked first in three benchmarks: RoboDojo, RoboTwin and RoboColiseum.
机器之心 · 微信公众号Source ↗Qiyuan releases PrimeSim world model simulator, topping EWMBench and WorldSimProbe
The Qiyuan Embodied Research Center released PrimeSim, a world model simulator for robot manipulation tasks, built by adapting a pretrained video generation model through mid-training and post-training stages. It ranked first on EWMBench with a composite score of 6.09 and took first place in three of five WorldSimProbe evaluations: global trajectory coverage, action-source behavior preservation and interaction dynamics.
机器之心 · 微信公众号Source ↗X-Era AI Tuoyuan Zhihui open-sources 4D world action model E0.5, with real-robot success rate surpassing π0.5
X-Era AI Tuoyuan Zhihui open-sourced the native 4D world action model E0.5, achieving overall average success rates of 85.0% and 88.3% on the AgileX Piper dual-arm robot and the Zerith H1 humanoid robot, 13.3 and 26.7 percentage points higher than π0.5, respectively.
机器之心 · 微信公众号Source ↗PokeBot releases embodied API RoboAPI, world-action model Poke-Aurora debuts
PokeBot released its first model product RoboAPI in the early hours of October 30, a publicly accessible, out-of-the-box embodied API; after developers connect and adapt a robotic arm, they need neither to collect data nor train models, and can execute tasks not specifically trained for by issuing instructions with a single prompt.
机器之心 · 微信公众号Source ↗Ropedia releases open world model W0, trained on about 10,000 hours of first-person human video
Ropedia released the open world model W0, trained on about 10,000 hours of first-person human video; a single set of weights can predict the first-person view of humans, robotic arms and humanoid robots (Unitree G1, Fourier GR-1), using a shared camera-plus-skeleton representation, and integrating a new robot requires only its URDF.
The Humanoid HubSource ↗AgiBot releases WITA-OMNI interaction model and Lingqu OS 1.0, to be fully open-sourced by year-end
AgiBot released the embodied omni-modal end-to-end interaction large model WITA-OMNI at its 2026 Innovation Day, ranking first in four benchmarks: Daily-Omni, TimeLens-Bench, EmbSpatialBench and MMAU, with an average response of about 600 milliseconds.
机器人前瞻 · 微信公众号Source ↗
Product Releases/Updates
RLinf v0.4 released: an upgraded open-source training framework for embodied reinforcement learning
RLinf, an open-source embodied reinforcement learning training framework developed by Tsinghua University, Infinigence AI together with the Beijing Zhongguancun Academy and other institutions, officially released v0.4, iterating in four directions: a Lego-style real-robot system, real-robot post-training algorithms, omni-modal model support and a domestic computing power foundation.
Xbotics具身智能实验室 · 微信公众号Source ↗NVIDIA expands Isaac platform, adding humanoid reference platform, Isaac OS and Isaac Lab 3.0
At GTC Berlin, NVIDIA announced an expansion of the Isaac open-source robotics development platform, adding the Isaac OS runtime, the officially released Isaac Lab 3.0 learning framework, and an open-source reference workflow based on DGX Station.
Industry News
Unitree's market value fluctuated sharply after listing, embodied companies advanced financing and mass-production delivery in parallel, and overseas humanoid robot partnerships and listing processes also saw new progress.
Unitree Technology launches STAR Market IPO subscription: issue price 150.80 yuan, total market value after issuance about 61 billion yuan
Unitree Technology launched its STAR Market IPO subscription on August 10, with subscription code 787836, an issue price of 150.80 yuan per share, corresponding to an issue price-to-earnings ratio of 219.23 times, a total market value after issuance of about 60.993 billion yuan, expected fundraising of about 6.099 billion yuan, and an oversubscription of about 45%. Unitree's 2025 revenue was about 1.699 billion yuan, net profit attributable to parent about 278 million yuan, and actual shipments exceeded 5,500 units; the article, drawing on precedents such as UBTech, Geek+ and Estun, points out that there is no general rule between robotics themes and first-day gains.
OFweek RoboticsSource ↗The Astra shockwave and the embodied industry's 50 days
In the 50 days after the release of GPT-6 Astra, the embodied industry carried out intensive evaluations and discussions around it, and general large models began to directly take over robot operation. Tests by SenseTime together with multiple institutions concluded that Astra has basically acquired the ability to complete embodied tasks, but still has an obvious gap from practical use in real environments, underperforming in millimeter-level contact, dynamic control and response speed. The article also reviews industry uses such as using Astra directly for policies, reviewing VLA actions, building simulation environments and data annotation, as well as chain reactions such as cooling financing and big companies stepping up data collection.
钛媒体Source ↗Unitree's market value retraces nearly 70%, Wang Xingxing admits at the Industry Expo that he did not believe enough in AI
Less than two months after Unitree Technology's listing, its share price fell from a high of 1,237 yuan to around 420 yuan, wiping out more than 270 billion yuan in market value, with a maximum drawdown of nearly 70%. Founder Wang Xingxing said on October 12 at the Shanghai Industry Expo that his biggest mistake over the past decade-plus was not believing enough in AI, and gave a prediction that the robot era could see an inflection point in as soon as one to two years; Unitree is building a physical AI self-evolution system, and its share price closed up 3.64% the day after the remarks.
OFweek RoboticsSource ↗Zhongke Huiling completes several hundred million yuan A++ round, advancing embodied intelligence mass-production delivery
Zhongke Huiling completed a several-hundred-million-yuan A++ financing round, jointly invested by Zhilai Investment, Yonghua Investment, Maotian Investment, Yuanhe Chenkun, Guoke Kezhi, Luning Investment and others, with funds used for product mass-production delivery, embodied brain technology R&D iteration and market expansion. The company's self-developed CASBOT PA perception-to-action model has been validated at actual workstations in a shoe factory in Putian, with the outsole stamping process reaching a cycle of 7 seconds per piece and sole pressing reaching 18 seconds per pair. In 2026, its total in-hand embodied shoemaking robot orders exceeded 400 million yuan, and equipment scheduling entered the thousands-of-units scale.
中科慧灵 · 微信公众号Source ↗NEURA Robotics releases humanoid robot 4NE1 Gen 4, and opens TUM RoboGym with the Technical University of Munich
NEURA Robotics and the Technical University of Munich (TUM) opened TUM RoboGym at LabCampus near Munich Airport, covering more than 2,300 square meters, co-built with TUM MIRMI, with an initial three-year cooperation; the two sides invested about 17 million euros in total, of which NEURA contributed about 11 million euros.
SEC declares effective the S-4 registration statement for the merger of Agility Robotics and Churchill Capital Corp XI
The Form S-4 registration statement for the proposed business combination between Agility Robotics and Churchill Capital Corp XI (Nasdaq: CCXI) was declared effective by the SEC on October 13, 2026, and Churchill XI is scheduled to hold an extraordinary general meeting on November 2, 2026 for a vote.
Dream Labs raises $90 million seed round, and launches Dream API and Dream-1 model
Dream Labs announced the completion of a $90 million seed round led by B Capital, with participation from Kindred Ventures and others. The company also launched Dream API, which can directly convert robot camera images and natural-language instructions into motor commands, and said its flagship model Dream-1 is trained on videos of everyday human tasks and can zero-shot control any robot with a gripper.
Robotics & Automation NewsSource ↗
Paper Research
Research focused on world model benchmarks and cross-embodiment control, with new frameworks and evaluation results released in directions including collective experience learning, online exploration and general tactile policies.
Infinigence AI with Tsinghua and AgiBot proposes collective experience learning CEL, with a 98.3% success rate validated on 96 robots
Wuwen Xinqiong, together with Tsinghua University and Zhiyuan, proposed the Collective Experience Learning (CEL) paradigm and completed prototype validation on 96 robots. In the phone packaging task, the policy trained with CEL completed 120 tests on two same-model robots that did not participate in training, with the success rate rising from 79.2% to 98.3%; when the number of robots participating in learning increased from 16 to 96, the test success rate rose from 86.7% to 98.3%.
3D视觉工坊 · 微信公众号Source ↗Harvard, NVIDIA and others propose RAX: a robot agent explores online, leading by an average of 37.4 points across five major benchmarks
Research teams from Harvard, NVIDIA, Berkeley, MIT and Stanford proposed RAX (Robot Agent eXploration), which connects a general coding agent to robot perception and control interfaces, and through online exploration writes successful decisions into executable programs and accumulates them as skills, while model parameters remain frozen.
CVer · 微信公众号Source ↗WorldOracle releases an action-conditioned world model benchmark; seven models show deviations between predictions and real execution
Princeton University, the University of Texas at Austin, the University of Toronto, NVIDIA and AgiBot jointly launched the action-conditioned world model benchmark WorldOracle, using the same initial frame and action sequence to compare model predictions with real execution.
The Robot ReportSource ↗Delft University of Technology's MonoRace achieves autonomous drone racing at 100 km/h using only a single front-facing camera
The MonoRace system from the MAVLab at Delft University of Technology in the Netherlands relies only on a single front-facing camera and an IMU, and won consecutive victories against three world-champion-level pilots at the 2025 Abu Dhabi Autonomous Drone Racing competition; the paper was published in Science Robotics on October 14, 2026.
深蓝AI · 微信公众号Source ↗X-WBC: a cross-embodiment foundation model for humanoid whole-body control
X-WBC is a cross-embodiment humanoid whole-body control foundation framework that separates relatively shared human motion semantics from embodiment-specific physical execution. It uses human-centric command tokens to align full human motion, robot reference motion and sparse VR observations, employs a causal Transformer to learn reusable temporal structure from mixed multi-robot rollouts, and a lightweight robot-specific module then maps the shared representation to each robot's proprioception and action space.
arXiv cs.ROSource ↗SenseTime and a joint team evaluate the embodied capabilities of GPT-6-Astra and open-source the SenseNova-RoboRSI framework and a 27B model
A joint team formed by SenseTime and the Hong Kong University of Science and Technology (Guangzhou) and others released the report "Has GPT-6-Astra Achieved Embodied Intelligence?", systematically evaluating the embodied capabilities of GPT-6-Astra on 12 benchmarks, and simultaneously open-sourced the SenseNova-RoboRSI evaluation framework and a 27B embodied model.
3D视觉工坊 · 微信公众号Source ↗RynnWorld-Latent: a cross-embodiment world model conditioned on general latent actions
The research team proposed RynnWorld-Latent, a cross-embodiment video world model that uses RynnLAM's 3D-grounded latent actions as a shared conditioning interface and can be jointly pretrained on human interaction videos, robot demonstrations and simulation data.
HuggingFace Daily PapersSource ↗Study uses an activation injection mechanism to analyze vision-language-action models
Researchers proposed an interpretability pipeline for analyzing what a fine-tuned vision-language-action (VLA) policy is computing internally. On pi0.5, injecting the activations of task A into the action expert while letting the policy view the scene of task B caused 99.6% of episodes (n=1968) to execute the motion trajectory of task A, indicating that the injected program is bound to absolute workspace coordinates and ignores the visible scene.
arXiv cs.ROSource ↗Rendering-free event-frame simulation accelerates reinforcement learning for autonomous robot navigation
The authors proposed an analytical event simulator that directly synthesizes event streams from scene geometry and ego-motion without image rendering, generating 64x64 event observations on a single GPU in only 17 microseconds, more than 400 times faster than advanced rendering pipelines. A reinforcement-learning obstacle-avoidance policy for event cameras trained with this simulator converges in under 30 minutes, three orders of magnitude faster than training with rendered events, and can transfer zero-shot to Isaac Sim ray-traced event simulation.
arXiv cs.ROSource ↗FTP-1: a general foundation tactile policy across tactile sensors
The research team proposed FTP-1, a general foundation tactile policy pretrained across multiple tactile sensors and embodiments, which uses heterogeneous encoders to map image, array and state-type tactile signals into unified morphology-aware latent tokens, jointly modeled by a shared tactile Transformer expert.
arXiv cs.ROSource ↗Evaluation of a frozen frontier VLM used directly as a robot policy on the RoboMME memory-dependent manipulation benchmark
The study uses the frozen frontier vision-language model GPT-6-Astra directly as a robot policy and tests it on the RoboMME benchmark containing 16 memory-dependent manipulation tasks; the model retains the entire episode context throughout and directly outputs each end-effector waypoint and gripper command, without any training specific to this benchmark.
arXiv cs.ROSource ↗SuperMimic: scaling humanoid whole-body control with a differentiable world model
SuperMimic uses a differentiable world model with shared dynamics to train the world model and the tracking policy simultaneously in a single training loop, providing analytical gradients to the policy through backpropagation-through-time over short imagined trajectories.
arXiv cs.ROSource ↗RepresentNav: a navigation foundation model built on a self-constructed representation space
The authors proposed the RepresentNav navigation foundation model, which uses Representation Residuals, a three-layer transformer, to aggregate full-depth features from multiple visual foundation models and construct a representation space driven by the policy's own objectives.
arXiv cs.ROSource ↗RoboCoach: using a world model to locate weak subtasks and improve long-horizon task success rates through targeted teaching
The team of Zhang Changshui and Cui Sen at Tsinghua University proposed the embodied intelligence system RoboCoach, which uses the world model CoachWorld to run long-horizon tasks in closed loop in imagination, locate the first timed-out weak subtask and request targeted human teaching, updating only the corresponding skill expert. Under a fixed demonstration budget, the Franka success rate rose from 13.3% to 75.0%, and AgileX from 40.0% to 83.8%, while the uniform-collection baseline reached only 30.0% and 47.5%, respectively. The average success rate on four unseen new task routes was 35.0%, versus 0% for the shared-policy baseline.
机器之心 · 微信公众号Source ↗