On October 22, at the main forum of the 2026 iFlytek Global 1024 Developer Festival, the device-cloud integrated iFlytek Spark X2.6 was released, with capabilities on both the cloud and device sides (Spark X2.6-4B) improved in sync, balancing task performance, security, and cost through precise division of labor and on-demand collaboration.
XPeng announced that its second-generation VLA large model, after an upgrade, has been deployed on-device to the IRON humanoid robot, giving it temporal understanding and future behavior prediction capabilities. IRON re-reasons every 0.2 seconds based on the scene before it and predicts motion changes within 6 seconds, allowing it to respond to last-minute changes during handshakes, fist bumps, and high-fives. XPeng plans to achieve mass production of humanoid robots by the end of 2026, begin deliveries in the first half of 2027, and first deploy them in commercial scenarios such as stores and shopping malls.
Google DeepMind demonstrated a robot training technique alongside the release of Gemini 4 Argon: after watching raw video of a person playing cornhole, the model analyzes the environment, maps spatial coordinates, and computes full-body joint motion, automatically generating simulation code, then transfers the learned behavior to Apptronik's Apollo humanoid robot, which hit seven consecutive shots in real-machine testing.
RockAI Matrix, under RockAI Technology, together with universities including the Hong Kong University of Science and Technology and Carnegie Mellon University, released and open-sourced EdgeRock-4B. The model is converted from Qwen3.5-4B, and its core is the Memory Rock Anchor (RMA) mechanism proposed by the team, which keeps the KV Cache from growing with context.
iFlytek has released the end-cloud integrated large model iFlytek Spark X2.6, comprising the cloud-side 293-A30B model and the already open-sourced device-side X2.6-4B model. The cloud model supports a one-million-token context, with an input price of 0.9 yuan per million tokens and an output price of 3.6 yuan per million tokens, lower than DeepSeek-V4.1-Flash-1017.
Dream Labs announced this week that it has raised a $90 million seed round and launched the Dream API, which converts robot camera feeds and natural-language instructions directly into motor commands. Its flagship model, Dream-1, is trained on videos of everyday human tasks; the company says it can control new gripper-equipped robots zero-shot without additional data. It currently supports tasks such as tabletop pick-and-place and clothes folding, with wheeled mobile robots and dexterous hands to be supported later.
Embodied intelligence evaluation platform RoboColiseum released an evaluation report: Claude Opus 5.5 ranked first among four multimodal models with 76.28 points, followed closely by GPT-6 Astra with 75.99 points, while GPT-6.1 Sol and Gemini 4 Argon scored 70.91 and 66.86 points respectively; the top two on the official overall leaderboard remain specialized embodied policy models Paimon (77.54 points) and HiDream-O1-Embodied (76.92 points).
At the 2026 Chery Global Innovation Conference, Chery released CheryMind, a vertical foundation model for physical AI. It serves as the technology foundation of the Yaoguang 2035 strategy, covering automotive R&D and manufacturing and extending to businesses including intelligent driving, intelligent cabins, robotics, eVTOL, and agricultural machinery.
NVIDIA has expanded the Cosmos platform with new training tools, simulation checkpoints and Cosmos Data Factory blueprints for training, evaluating and deploying world models.
Why it matters: The original lists Cosmos's new training tools, simulation checkpoints and data blueprints, and gives specific data on users such as Porsche and ANYbotics, which can be used to assess changes in the world model development workflow.
SEER Robotics demonstrated its robotic arm completing the full coffee-making process without human intervention, from picking up the portafilter, receiving grounds, distributing and tamping the grounds, locking the portafilter, extraction, to cleaning. This is the longest multi-step robotic coffee task the company has demonstrated to date.
Ken Goldberg shared the work of his postdoc Shuangyu0_0Xie: she found that Gemini 4 Argon performed well on spatial robotics queries, so she used Claude to reverse-engineer its skill library and implemented the robotics harness described in the Gemini 4 Argon report. The harness enabled Gemini to complete a zero-shot bimanual handover, and it is now open source.
SuperMimic trains a humanoid full-body control policy on 129,514 motions, surpassing the previous best result of 128 GPUs over 7 days with 1 GPU in 3 hours, reducing GPU hours by over 1,000x. Pieter Abbeel shared it and called it a game changer, noting that the key lies in using a world model faster than conventional simulation and in collecting far more signal than policy gradients through backpropagation through time. The project page is https://supermimic.github.io, and the demo is at https://supermimic.github.io/demo/.
ACE Robotics released ACE-Brain-0.6, a vision-spatial-anchoring-enhanced embodied foundation model focused on Grounding capability. It surpasses GPT-6 Astra on six of eight Grounding benchmarks, with an average score of 74.53 across seven public benchmarks (Astra: 73.68), and scores 94.13 on its self-built real-robot evaluation set RobotGrounding (2,591 questions, 14 task types; Astra: 91.10).
NVIDIA released and open-sourced the data engine Flora, which converts raw robot video into training data such as timestamped action descriptions, process rewards, and outcome rewards; it is already used to train the GR00T series models. Flora consists of an action description model, a temporal segmentation model, and a reward model, fine-tuned from the Qwen3.5 series of vision-language models, and can label 200,000 hours of data per day on 512 GB300 GPUs.
Why it matters: NVIDIA's open-source data engine cuts robot video labeling costs to about 1/125 of a general-purpose large model and provides several benchmark comparisons, which can be used to assess the substitution space in the data labeling segment.
NVIDIA has released and open-sourced Cosmos3-H-Dreams, a real-time generative surgical world model that can generate surgical imagery in real time based on robot operation commands, simulating instrument movement, tissue traction, and deformation, with support for laparoscopic cholecystectomy scenarios.
NVIDIA and the UW Medicine Institute for Protein Design launched the open model RFD4-Proteina, combining programmable protein design with search capabilities. In four benchmark tests reported in a research draft—protein binding, small-molecule binding, enzyme scaffolds, and DNA binding—computational design success rates and generation throughput were both higher than the evaluated baselines.
Nvidia announced an expansion of its open model lineup, introducing the Nemotron Retriever retrieval model, the Kumo structured enterprise data model, the Cosmos 3.5 world model, the Alpamayo 2 Edge autonomous driving model, the BioNeMo protein design model, and the Ising quantum computing model, among others. In a benchmark covering proteins of 200 amino acids, BioNeMo RFD4-Proteina generated designs about 13 times faster than RFdiffusion3.
Tsinghua-affiliated embodied intelligence company Qianjue Technology released the predictive world model Polibrain at the 2026 Physical AI and World Models Conference. It focuses on learning physical laws and cross-embodiment reuse; rather than generating complete future frames, it predicts state changes in a task-relevant latent representation space, forming a closed loop of perception, decision-making, execution, and feedback.
It turns out that getting a few important details right is all it takes: extracting policies using only advantage-based supervised learning can achieve state-of-the-art results in reinforcement learning (RL)!
Thinker-WM is a unified multimodal world model that jointly models and generates RGB, metric depth, and robot actions through a shared Unified Dynamics Transformer; the same set of parameters supports joint video-action generation, world prediction, policy inference, action-conditioned generation, video-to-action generation, and inverse dynamics.
The research team released ACE-Brain-0.6, which uses a unified data pipeline to combine robot manipulation recordings with complementary visual data from multiple sources, automatically annotates them, and converts the result into verifiable grounding supervision; the model was obtained by fine-tuning Qwen3.5-9B on approximately 1.5 million samples. ACE-Brain-0.6 outperforms GPT-6 Astra on four of seven public grounding benchmarks, with an overall average score of 75.41 versus 73.68; under a fixed fine-tuning setup, overall grounding performance improves steadily as the dataset expands to cover more grounding scenarios.
Under the Gravity framework, Geek+ embodied intelligence proposes the Gravity Control causal intervention model. Without training the original policy model or adding new demonstration data, it uses mechanistic interpretability to uncover the model's internal action regulation mechanisms and intervenes on relevant latent features at inference time to achieve runtime correction.
Genrobot.AI reposted and congratulated Dream Labs on launching Dream-1, saying it is a launch partner for Dream-1. The quoted content says Dream Labs has exited stealth and released Dream-1, a robot foundation model that can control dual-arm robots it has never seen before, complete tasks that were not demonstrated, and work in scenes it has not visited, at a price of 3 USD per hour, with the Dream API opening to everyone that day.
Nature Will recently released a demonstration video of the Natura-1 model, in which robotic hands, without human assistance, completed tasks such as wrapping wontons, packing express deliveries, brewing tea, applying phone screen protectors, and writing Chinese calligraphy with a brush. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model, with data processing covering nearly 1 million hours of first-person human interaction data. It can drive a 20-degree-of-freedom dexterous hand within a 100-millisecond control loop and follow complex natural language instructions to generalize in unseen scenarios.
Dream Labs released Dream-1 yesterday. A Reddit user who deployed it on a robotic arm said it is fast and can perform tasks zero-shot, but it also exhibits random behavior, such as repeatedly grabbing a red block when asked to move an apple. The model has been made public for people to try.
Cubic Doggo used Gymnasium and SB3's PPO in MuJoCo to train standing while minimizing pitch and roll, and deployed the ONNX file back to MuJoCo, raising resistance to horizontal push from 3.5N to 4N. The second phase added "push recovery training," using horizontal force as domain randomization under a curriculum schedule, but after two weeks of training the improvement was limited. Next steps are balancing on fixed and varying slopes, then deployment to Gazebo and a physical robot.
Dream Labs released Dream-1 yesterday. A user deployed it on a robotic arm, saying it is very fast and can perform tasks zero-shot. In testing, when asked to move an apple, it instead grabbed a red block and did not stop, showing occasional random behavior. The publisher has made Dream-1 publicly available for people to try.
VAST has open-sourced EdenLive, a real-time generative renderer in the Project Eden series. The upstream interaction system maintains game rules and state, while EdenLive handles real-time image rendering. It supports white-box models, 2D/3D scenes, or real footage as references, and changing the first frame and prompt can alter the look among photorealistic, hand-drawn, cartoon, and other styles.
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; adding a new robot requires only its URDF.
Why it matters: The original article provides the training data scale, the cross-embodiment transfer method, and measured comparisons, so readers can judge the world model's actual impact on robot policy training workflows.
VAST has open-sourced EdenLive, a real-time generative renderer in the Project Eden series, including 5B and 14B model weights (each with bidirectional generation and few-step streaming generation versions), real-time streaming inference code, EdenGolem v0.1, a front-end workbench, and a technical report.
The Xingchen General Artificial Intelligence Laboratory of China Telecom Artificial Intelligence Technology Co., Ltd. has released and open-sourced Xing4.0-29B-Intent, a model dedicated to intent understanding, and has simultaneously opened 2.753 million supervised fine-tuning training corpus entries, model weights, and inference images, with the code licensed under Apache 2.0.
Nanjing University, JD.com and other institutions have open-sourced VideoJev, which unifies video temporal grounding and multiple-choice video question answering into a candidate scoring task, enabling multimodal large models to directly output structured judgments.
MemT-1.0 releases the MemT-1.0-Nano and MemT-1.0-Flash memory foundation model family, internalizing memory as part of model computation. Built on Metis, it employs Gated DeltaNet-2 and Mixture-of-Storage.
MemTensor has released and open-sourced the memory-native MemT 1.0 series, comprising two models, MemT-1.0-Flash and MemT-1.0-Nano, with parameter counts of 288B and 36B and activated parameters of 17B and 3B, respectively.
A joint research team from SenseTime, Nanyang Technological University and other institutions has introduced the SenseNova-RoboRSI unified Harness, which raises GPT-6-Astra's manipulation success rate on 42 RoboDojo tasks from 15.83% in the baseline configuration to 54.17%, with model parameters unchanged.
Zymbiotek created a tabletop battle simulation in MuJoCo of two open-source Jumper robot crabs for a KingKong Robotics competition. The motions are physics-simulated based on measured servo torque curves, and the dance is a reinforcement learning policy that can run on the real robot. The counter uppercut is a real simulated motion, with only the final flip being animation—testing showed one Jumper cannot flip another.
Dream Labs has launched Dream-1, a robot foundation model that can control never-before-seen dual-arm robots, complete tasks that were not demonstrated, and adapt to unfamiliar environments. The Dream API opened to everyone the same day, starting at $3/hour. GenRobot said it provided human data collaboration for Dream-1 and called this another breakthrough in zero-shot research.
Researchers from XPeng and Peking University, the University of Hong Kong, UC Berkeley, Princeton University, National University of Singapore, Tsinghua University, and Hong Kong University of Science and Technology (Guangzhou) jointly proposed a 4B perception foundation model, GroundingPI, which unifies point selection, boxes, text, document regions, and interface controls into a single spatial token output protocol. The architecture uses a MoonViT-V2 vision encoder and a Qwen3-4B language backbone, and training includes supervised fine-tuning and GRPO reinforcement post-training.
Singapore-based embodied intelligence startup Ropedia, together with Nanyang Technological University, Duke University, UC San Diego, and Tsinghua University, released the cross-embodiment mobile manipulation world model Ropedia W0, which uses a unified action representation to predict first-person future scenes after humans, robotic arms, and humanoid robots perform actions.
The SenseTime SenseNova team launched the SenseNova-RoboRSI unified execution framework and the 27B embodied model SenseNova-RoboRSI-27B-Preview.
Why it matters: Through comparison data from the same model with different execution systems, readers can intuitively understand the magnitude of the execution framework's impact on embodied task performance and its cost.