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.
AgiBot has released its embodied Agent framework Genie Master. A Genie G2 running the framework, when deployed in a real convenience store, does not rely on human demonstrations or debugging; it can identify the causes of failures on its own, adjust strategies, and accumulate experience, improving continuously on tasks such as packaged-food packing, picking, and restocking. AgiBot says this is the industry's first validation of Physical RSI (embodied recursive self-improvement) outside simulation in a real business scenario, and it plans to arrange 6 Genie G2 units on November 2 for a continuous 24-hour livestreamed convenience-store operation test.
Official weights of InsertAnything's force-conditioned insertion policy were evaluated in Isaac Lab: at 0.1 mm clearance, baseline success rates for two hole types were 96–100%. With a 5 mm initial hand offset, success rates were unaffected, but with a 3 mm hole-position observation deviation, the hexagonal hole success rate dropped to 60%.
Xingyan Intelligence has launched Auriga, a general embodied Harness framework. Under a zero-shot protocol that involves no task-specific pre-tuning, no reliance on expert demonstrations, single-attempt execution, and no environment resets, it achieved a 57.74% average task success rate and an average process score of 64.00 across 42 tasks and 6,300 test episodes on RoboDojo, using GPT-6 Astra as the base model. It leads the leaderboard on both metrics.
Fysics AI, together with Fudan University's CITLab, proposed the unified world-action model Fysiverse-UniWAM, which uses a single model to handle autonomous navigation, mobile manipulation, and fixed-workstation dexterous manipulation.
Why it matters: Readers can see a unified world-action model approach that does not rely on large-scale robot pretraining, along with the specific design of a shared backbone for navigation and manipulation.
Tencent Hunyuan used the open-source world model Hy-Nomas to reproduce 8 demos from the official Google Genie 4 website and released a side-by-side comparison; it is close in interactivity and content completeness, but still lags in native audio, camera smoothness, and visual style consistency.
Google DeepMind released the Genie 4 world model research preview on October 27, marking the first native support for synchronized audio-video generation, extending the duration of a single coherent world from 1 minute in the previous generation to about 2 minutes, and supporting real-time scene modification via typing. Google also demonstrated an experiment in which Genie 4 worked with Gemini 4 Argon to direct a humanoid robot in playing checkers, but it is currently available only to a small number of testers, and ordinary users need a Google AI Ultra subscription.
Qiyuan Robotics launched PrimeSim, a world-model simulator for robot training, on October 30 and topped the WorldSimProbe LIBERO leaderboard. The simulator builds an interactive, trial-and-error-friendly, continuously learning virtual world for robots, used for policy validation and reinforcement learning. In testing, reinforcement learning with it raised policy success rates from 61.7% to 73.2%.
WorldOracle presents a case where the real gripper failed to grasp the fruit, but the generated rollout showed the fruit being picked up, despite the model receiving the same initial observation and recorded actions. When evaluating predictions, check separately whether the robotic arm follows the trajectory and whether the object's final state is correct; for an empty grasp, predicting an empty grasp is the correct answer, while a seemingly successful grasp video may actually be wrong.
On October 30, Qiyuan Robotics launched PrimeSim, a world-model simulator for robot training, and topped the WorldSimProbe LIBERO leaderboard. The simulator builds an interactive, trial-and-error, continuously learning virtual world for robots, supporting policy validation and reinforcement learning. In testing, reinforcement learning with it raised policy success rate from 61.7% to 73.2%.
Humans should be more coordinated than robots. But that's not the case at SIGRobotics. @_robert_molina trained one of our robotic arms with reinforcement learning to bounce a ping-pong ball on a paddle continuously for up to 7 minutes. Watch this time-lapse video: it just keeps bouncing.
At its 2026 Innovation Day, AgiBot released WITA-OMNI, an embodied omni-modal end-to-end interaction large model. It ranked first on four benchmarks: Daily-Omni, TimeLens-Bench, EmbSpatialBench, and MMAU, with an average response time of about 600 milliseconds.
Why it matters: AgiBot released an interaction model, a native operating system, and an ecosystem plan all at once, giving readers a way to understand its path to opening up its technology foundation and how it accumulates real-world interaction data.
Google DeepMind has released Genie 4, a new-generation world model that supports real-time generation of explorable virtual worlds, modification of weather and scenes, simulation of physical effects such as collisions and gravity, and simultaneous generation of ambient sound effects and character dialogue. Genie 4 is Google's first real-time world model to support native audio generation; it is also integrated with Gemini 3.8 Flash as a real-time guiding agent, with Gemini understanding goals, observing the environment, and deciding the next action.
Daxiao Robotics, in collaboration with Nanyang Technological University's S-Lab, has released Mundo, a world model that reconstructs 1 to 12 real-scene images into an explorable, physically interactive 3D simulation environment. The model has built-in physical properties, with MuJoCo computing object motion under gravity and contact constraints, and it supports training-data generation and policy testing for tasks such as navigation and pick-and-place. The team has reconstructed a hotel floor and laundry area for training service robots on navigation and interaction tasks.
At the 2026 China Marine Economy Expo, Lead Robot released "Dingfengbo," a world model for shipbuilding. It takes design drawings, process documents, and bills of materials as unified inputs, parses the engineering logic among them, breaks an entire ship down into blocks, sub-assemblies, sub-assemblies and parts, and reconstructs the 3D structure. This frees robots from reliance on manual teaching, enabling them to perceive autonomously, decompose tasks, and decide how to work.
The Qiyuan Embodied Research Center has 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 ranks first on EWMBench with an overall score of 6.09, and took first place in three of five WorldSimProbe evaluations: global trajectory coverage, action-source behavior preservation, and interaction dynamics.
Why it matters: The original text provides the two-stage training recipe and results across multiple leaderboards, which readers can use to judge the feasibility of adapting video generation models into simulators.
PokeBot released its first model product, RoboAPI, in the early hours of October 30. It is 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 issue commands with a single prompt to execute tasks the model was not specifically trained on.
Why it matters: Readers can learn how a publicly accessible embodied API hands zero-shot generalization capability over to external developers for testing, as well as its success-rate data from its own evaluation.
Daxiao Robotics, in collaboration with Nanyang Technological University's S-Lab, has released Mundo, an interactive world model built for specific environments that integrates 3D scene reconstruction, physics simulation, and video generation. The system uses scene code to record object weight, center of mass, and friction coefficients, computes motion and collisions with a 960Hz MuJoCo physics engine, and then renders visuals combined with on-site photos; inputting 1 to 12 phone photos is enough to reconstruct a 3D space that supports pushing, pulling, and collisions.
AgiBot has proposed a robot workload measurement framework called Worken, which multiplies three factors—the number of active robots, autonomous working hours, and output efficiency ratio—to estimate the effective intelligent work hours completed by robots in real deployments. The framework was proposed by co-founder, president, and CTO Peng Zhihui, and emphasizes distinguishing powered-on time from autonomous operating time, and measuring qualified output against task benchmarks.
Xingyuan Intelligence has released Chorus, a cross-embodiment data engine for robot pretraining that can convert human manipulation data collected by devices such as smart glasses, chest-mounted cameras, wrist cameras, and UMI into training data adapted to different robot embodiments.
Qiyuan Robotics has launched PrimeSim, a world-model simulator for robot training, and topped the WorldSimProbe LIBERO leaderboard. PrimeSim builds an interactive, trial-and-error, continuously learning virtual world for robots, used for policy validation and reinforcement learning. In tests, reinforcement learning with it raised policy success rate from 61.7% to 73.2%.
X-Era AI Tuoyuan Zhihui released the open-source native 4D world action model E0.5, which uses a 3D foundation model as its backbone and integrates future 3D geometry prediction directly into action learning. On the AgileX Piper dual-arm robot and the Zerith H1 humanoid robot, after pretraining on about 160,000 robot trajectories, E0.5 achieved overall average success rates of 85.0% and 88.3%, respectively, improving on π0.5 by 13.3 and 26.7 percentage points; its out-of-distribution average success rates were 73.3% and 76.7%.
Why it matters: The original text provides a comparison of E0.5 and π0.5 success rates on out-of-distribution real-robot tasks, which readers can use to judge the effect of 3D geometry supervision on manipulation generalization.
Agibot launches the embodied model GO-3 Preview, built on an embodied-native WRAM (World Reasoning Action Model) architecture that connects task reasoning, future prediction, and action generation, supporting full-pipeline scaled learning across pretraining, post-training, and post-deployment.
AgiBot released the GO-3 Preview embodied foundation model and plans to open-source it by the end of the year. In the same period, AgiBot disclosed first-half revenue of nearly 2 billion yuan and introduced a "effective intelligent working hours" metric for robots.
At the 2026 Agibot Innovation Day, Agibot released its next-generation embodied foundation model GO-3 Preview and plans to fully open source the model and the embodied Agent framework Genie Master by the end of 2026. On the same day, Agibot also released Worken, the first "work-hour KPI," introducing KPI assessment standards for robots.
The PhAIL team and Nebius have launched a robot policy competition. Participants package their policies as Docker submissions, which are first scored in two simulations: MolmoSpaces and RoboLab starting November 16. In December, the top 10 will run the same image on a real Franka. The tasks resemble the DROID benchmark but use a different dataset.
AgiBot has released WITA-Omni 1.0, an omni-modal interaction model that supports full-duplex dialogue with simultaneous speaking and listening, runs on-device with an end-to-end interaction latency of about 400ms, and integrates vision, speech, real-time dialogue, facial expressions, and physical actions into one continuous system. It currently ranks highly on multiple interaction benchmarks.
Google DeepMind has released Genie 4, a new-generation world model that extends single-session world coherence from 1 minute in the previous generation, Genie 3, to 2 minutes, and becomes Google's first real-time world model with native support for synchronized audio-video generation.
Daxiao Robotics has released the ACE-Brain-0.6 API, opening its spatial grounding capability to external developers through the ACE Hub platform, so they can call it without deploying the model independently. A simultaneous upgrade expanded high-quality spatial grounding training data from about 1.1 million to 1.5 million entries, raised the model's average score on seven public grounding benchmarks to 75.41, and increased inference speed beyond that of GPT-6 Astra.
At its Innovation Day, AgiBot released the GO-3 Preview embodied foundation model, the WITA-OMNI native multimodal interaction model and the AGILE general motion intelligence algorithm framework, and also launched Lingqu OS 1.0 and the AHA3 third-generation hardware architecture.
Why it matters: AgiBot released three types of models—task execution, interaction and motion—along with software and hardware architectures in one go, giving readers a clear view of its deployment approach across the full physical AI chain.
The author made a caniuse-style comparison table showing which robot learning policies can run on which robots, covering 24 robots and 21 policies (including ACT, Diffusion Policy, SmolVLA, π0/π0.5, GR00T N1.7, X-VLA, OpenVLA-OFT, and others).
At its Innovation Day, AgiBot released the GO-3 Preview embodied foundation model, the WITA-OMNI native multimodal interaction model, the AGILE general motion intelligence algorithm framework, the third-generation hardware architecture AHA3, and Lingqu OS 1.0, covering the three intelligence lines of operation, interaction, and motion.
AgiBot has released the embodied model GO-3 Preview, focused on full-pipeline scaled learning across pretraining, post-training, and post-deployment, and it ranked first in three evaluations: RoboDojo, RoboTwin, and RoboColiseum.
Why it matters: AgiBot disclosed GO-3 Preview's three-stage learning pipeline of pretraining, post-training, and post-deployment, along with specific benefit data from mixed training using real-machine clusters and world models.
X-Era AI has open-sourced its 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, respectively, 13.3 and 26.7 percentage points higher than π0.5.
Why it matters: The original text provides real-robot success rate comparisons between E0.5 and π0.5 on dual-arm and humanoid platforms, which can be used to assess the manipulation generalization performance of the 3D geometric backbone approach.
Under its Gravity embodied intelligence framework, Geek+ has launched the Gravity Control causal intervention model, which can correct local action deviations in robots with a single natural-language prompt, without updating the original policy weights or adding new demonstration data.
AgiBot officially released its GO-3 Preview embodied model on October 29, taking Top 1 on robotics leaderboards including RoboDojo, RoboTwin, and RoboColiseum, and expects to fully open source it by year-end.
Why it matters: AgiBot links pretraining, post-training, and post-deployment learning into a single pipeline, which readers can use to compare differences in scaling approaches for embodied models.
On October 29, AgiBot released its next-generation embodied foundation model GO-3 Preview, which uses the WRAM (World Reasoning Action Model) architecture to connect task reasoning, future prediction, and action generation. It has ranked first on three major evaluation leaderboards: RoboDojo, RoboTwin, and RoboColiseum. AgiBot plans to fully open-source it by the end of 2026.
At IROS 2026, held September 27 to October 1 in Pittsburgh, USA, Daimon Robotics presented its full-chain physical intelligence system spanning perception, data, and models under the theme From Touch to Intelligence, centered on the Daimon-TWM tactile-anchored world model.
Apex Intelligence has released and open-sourced its first large model, Apex-35B-V1.0, with 35 billion total parameters, about 3 billion parameters activated per token, and support for a 262,000-token context; it is based on Qwen3.6-35B-A3B and optimized through mid-training and post-training.
At the 2026 AgiBot Innovation Day, Yao Maoqing, AgiBot co-founder, co-president, and president of the embodied business system, officially released the new-generation embodied foundation model GO-3 Preview. AgiBot plans to fully open source GO-3 Preview and the embodied Agent framework Genie Master by the end of 2026.