After OpenAI released GPT-6 Astra, it can control a robotic arm to complete grasping and assembly; Robocurve tests show a 95% zero-shot success rate on the block-into-bowl task, while Claude Fable 5.1 scored 40% over the same period.
At a roundtable discussion on October 12 at the 26th China International Industry Fair, Sharpa co-founder Xiang Shaoqing proposed a data pyramid framework in which human operation data, simulation data, and real-machine data build on one another. Unitree founder Wang Xingxing responded that he is more focused on the last few millimeters of manipulation precision in real-machine data, and Academician Wang Jian called for opening up industrial data. The article reviews the differences between Figure Index crowdsourcing, data collection at Tesla's Optimus Academy factory, and China's precision-first approach, and notes that the key dividing line is who first completes the full data loop from collection to use.
Newton has released the v1.7.0 feature version, expanding simulation diagnostics, camera rendering, viewer interaction, and rigid/deformable body coupling workflows. Experimental features include composable solver diagnostic output, GRU actuator driving (with ONNX checkpoint support), and particle-rigid body attachment; mimic joints now cover the Featherstone, SemiImplicit, XPBD, MuJoCo, and VBD solvers.
David1726 released the first public version of Embodied Agent Harness 0.6.0, providing two Python entry points to implement agents and skills, with built-in Panda insertion, nut placement, stacking, and LIBERO placement examples, and support for connecting external inference workers through a UTF-8 JSON process interface.
The article reviews the approaches of 9 organizations addressing the shortfalls in first-person (Ego) data collection: Sharpa uses an EMG wristband moved forward as an intent entry point to fill visual blind spots, Xspark AI released the HumanTouch tactile multimodal system and a 1,000-hour tactile dataset, Wuji Technology open-sourced MINT and EgoPipeline and released 1,021 hours of first-person annotated data, and Daxiao Robotics open-sourced ACE-Ego-Hand and produced about 5,000 hours of training data.
YuanCe Future recently released a humanoid foundation model system centered on Whole-Body Intelligence (WBI), split into two layers: the motion control foundation model S0 (cerebellum) and the intelligence foundation model S1 (brain).
The author integrated the open-source vision-language navigation model LightNav-0 into the Microduck simulation (MuJoCo). The robot receives instructions such as "walk to the wooden door frame and stop," predicts a path from first-person frames, and executes it via MPC and the official ONNX gait policy.
Hesai Technology has launched the 3D world data engine Kosmo, comprising the 8K ultra-high-definition AI spatial camera Kosmo One and the 3D content creation and management platform Kosmo World. The system can scan and reconstruct scenes such as rooms and buildings into editable 3D data for robot training and world model development.
On October 14, Poxiao Intelligence and Tinno Mobile launched SeePhys Hub, a full-body, full-modality data collection system. It consists of a four-camera head ring, two wristbands, three waist-and-leg motion trackers, and a tactile glove, and can synchronously collect first-person vision, hand pose and high-density tactile data, and full-body motion information.
On October 14, Hesai launched the Kosmo One spatial camera and the Kosmo World 3D content platform, covering spatial capture, 3D reconstruction, scene editing, and robot training workflows.
Coding agents (from @claudeai, @OpenAI, @GoogleDeepMind) are opening new doors in robotics, but how can their code be tested quickly and iterated into better policies? Agentic Real2Sim answers: use RoboSimStudio to evaluate and improve the Push Letter Agent. (The main post comments on this: "Another exciting result in Agentic Robotics!")
Registration is open for the Xbotics × Motphys Fifth "Reinforcement Learning and Simulation Software" online training camp. The camp lasts eight weeks and is free of charge, and participants with zero reinforcement learning background may also join. All courses and training are hosted on the Qiaogong cloud, requiring only a browser locally.
Auto Pixel analyzes XPeng IRON's commercialization path: IRON plans to reach mass production by the end of 2026 and begin deliveries in the first half of 2027, entering the European market in the second half of 2027; the Pengxing robotics business has raised over $900 million, with a post-investment valuation exceeding $6.3 billion. The article notes that guided tours are the main initial scenario, and external customers need reasons to pay for redundant configurations such as dexterous hands and 2,250 TOPS of computing power, while store deployment is only the starting point of a data loop; the claim that about 70% of automotive technology reserves can be reused is hard-pressed to cover core barriers such as the brain and dexterous hands.
IsaacLab Browser Demo has released version v4, adding edge-latched viewport joysticks and a bounded MicroDuck warehouse scene. Pull a stick back or press Escape to stop; turning in place uses a 0.25 m/s forward assist.
Isaac Lab has released Browser Demo v5, adding a MicroDuck warehouse interaction scene with convex colliders derived from enabled scene geometry, compound rack/pallet components, and 12 movable floor props.
Isaac Lab releases interactive browser demos v6, providing prebuilt browser simulations for the documentation. It adds a 24-slab domino cascade, VDC cable bending, and an Ant walking demo with steering and direct perturbations, while retaining the directories from earlier versions.
On October 12 at the Shanghai Industrial Fair, HKUST, Estun, and Freedo Technology reached a long-term strategic cooperation to jointly carry out "industrial physical AI foundation technology research." The cooperation focuses on industrial-scenario data generation, physical-law modeling, dynamic-environment simulation, and robot task decision-making, connecting the technology chain of spatial-scene generation, industrial-environment semantic understanding, high-fidelity physical simulation, model training, and standardized evaluation. The three parties will build a closed technology loop integrating 3D spatial perception, factory digital twins, virtual simulation training, and physical robot execution to help flexible manufacturing land.
MuJoCo visualization tool mjswan has released v0.11.7. Main changes include: moving multithreaded MuJoCo out of a single-threaded SPA, adding CommandJointPositionActionCfg to drive joints via commands, and drawing a box's cube-texture faces the way MuJoCo does. The full changelog covers v0.11.6 through v0.11.7.
ttktjmt released mjswan v0.11.8. This update keeps mesh colliders in their compiled state when drawing meshes in the Y-up direction, and includes one major change (#166). The full update range is from v0.11.7 to v0.11.8, and it was automatically released by github-actions on October 14.
The RoboSimStudio toolkit has been officially released, an open interactive toolkit for Agentic Robotics that helps agents create Real2Sim models with detailed dynamics. The cited content mentions that given a phone video of a robot wiping, Claude and GPT each built a full-dynamics digital twin within an hour, and the toolkit supports the Fable, Astra, and Argon agents in modeling sponges, cables, and cloth.
This tutorial derives the Jacobian-transpose torque mapping and the task-space spring-damper law from the principle of virtual work, implements end-effector impedance control for the Panda arm in MuJoCo, and measures the quantitative effects of stiffness, damping, and anisotropic stiffness ellipsoids on compliance.
GreyOrange acquired AI digital twin company Site Bionics last week; the transaction amount was not disclosed. Site Bionics was founded in 2023 by Adam Blair, former founding CTO of RADAR, and Don Gillett, former chief architect at Amazon Go.
The Galaxea LEMO dual-arm development platform is officially on sale, with a launch price of 49,999 yuan. The platform integrates dual A1Z seven-degree-of-freedom robotic arms, G1Z grippers, a multi-view synchronized vision system, and a CCU controller. It supports 14-DOF dual-arm coordination, a 3 kg rated payload (5 kg peak, excluding end effectors), and ±0.1 mm repeat positioning accuracy, and can connect to external compute such as Orin, PC, and Mac.
A Mecanum-wheel omnidirectional robot can be launched in Gazebo simulation with a single command. The article also explains how to solve the common problem of the wheels not turning in simulation. This is the first post in the tutorial series.
MuJoCo 3.16.0 was released on October 14, 2026. Frames gain scale, scaling, and massscale attributes that can scale their contents at compile time without modifying mjSpec. New savecompiled and savecanonical compile attributes preserve the original modeling style, and the core library, simulate, plugins, and examples can now be built with Bazel.
Why it matters: This release makes compile-time scaling and save round-trip consistency explicit capabilities, so you don't have to worry about losing original editing information when changing model specifications.
armature-ai-labs' indieduck project has released R25 simulation study v1, including a project hub and interactive assembly guide, showing the R25 CAD and re-authoring assembly steps for the R25 head base, jaw, and exterior.
PhysiClaw released hardware v0.43 build files for assembling the PhysiClaw device, with English and Chinese materials. Contents include a full assembly manual (HTML + PDF with exploded-view SVGs), a sourcing guide with extrusion technical drawings, technical drawings for the 7 frame extrusions, and STEP files for the 9 custom 3D-printed parts (recommended to print in black PA12 via SLS/MJF) plus a bilingual print guide.
nematode v0.7.0 (Phase 8) is released. Experiments show that the nematode connectome's learning advantage comes mainly from gap-junction placement rather than connection strength, and that this advantage still holds in a 12-segment kinematic body. The worm learns to weathervane toward a thermal set point but does not learn klinokinesis.
BAAI, together with Inverse Matrix, Peking University, and vLLM, released PhysisServe, described as the first physical AI inference system that balances intelligence and high-frequency throughput, enabling world models with tens of billions of parameters to output in real time at kilohertz frequencies, one to two orders of magnitude faster than existing frontier solutions.
On the RDK X5 development board, the official stereo depth estimation model is only StereoNet (IGEV/GRU architecture); there is no smaller ready-made network. The available options are compiled versions such as V2.4_int8 (640x352x3x2, 23fps) and V2.4_int16 (15fps).
On October 14, Poxiao Intelligence and Tinno Mobile jointly launched the full-body multimodal data collection system SeePhys Hub, consisting of a four-eye headband, two wrist bands, three waist and leg motion trackers, and optional tactile gloves, capable of synchronously recording first-person vision, hand poses, contact pressure, and full-body posture and force estimation.
Newton has released v1.7.0rc2, the second release candidate for Newton 1.7, expanding solver diagnostics, camera rendering, viewer interaction, and rigid/deformable body coupling workflows. The MPM material scaling fix changes existing tuning; parameters such as mpm.young_modulus must be halved and retested. URDF damping and USD unit compatibility modes must be adjusted according to the migration notes. The simulation stack is upgraded to MuJoCo 3.15, and CUDA 12 GPU support is deprecated in this version.
The UCLA RoMeLa lab has published a PhoneBot paper on arXiv: a bipedal robot 483 mm tall, weighing 1.8 kg, with 13 degrees of freedom and about $400 in material costs. A phone mounted on the torso serves as sensor, onboard computer, and interaction interface, handling motion control, image recognition, and voice interaction; speech-to-text and large-model inference still rely on an external laptop, and training uses simulation plus left-right leg mirroring. With a used phone, the total cost is under $500. The team describes it as a lab test device with no arms that does not perform complex tasks such as carrying loads or climbing stairs.
Huanwang Technology and Pyromind have co-built an AITV intelligent solution, performing scenario post-training on Qwen3-VL-4B. Focus recognition accuracy rose from 45% to 99.6%, the success rate of "see it, say it" reached 99%, and the success rate of "search and play" exceeded 90%. It has been deployed in "Xiaoyou Casting - Movie Search" and covers TV apps including Cloud Video Aurora, Kiwi, Mango, Youku, and Bilibili.
LunCoSim has released a nightly test build dated 2026-10-14, with installers for Windows x86_64, macOS Apple Silicon, macOS Intel, and Linux x86_64. A built-in auto-updater can check for and download the latest nightly build, and users can have an AI agent read the bundled documentation, skills, and examples to build and validate missions. The official notes state that these unsigned nightly builds are intended for testing only.
Pantheon's dataset audit released last month found issues in several popular open-source robotics datasets, including frozen or disconnected wrist cameras, gripper actions recorded about two frames earlier than the actual motion, recording stopping mid-action, and long idle tails after task completion; the RoboDrop paper (arXiv 2609.10021) also points out that execution errors, sensor drift, and timestamp misalignment are common errors.
Humaan Studio 0.0.1-alpha.1 extends the workflow from robot assembly and simulation to training, policy management, and physical telemetry. New features include running versioned MuJoCo training experiments locally or over SSH, configuring RunPod compute and monitoring jobs, mirroring robot joint positions in 3D through a read-only Live Link telemetry bridge, and policy registration with compatibility checks.
The author integrated real Ouster OS0 lidar data and a fixed iPhone camera into a self-built "robot dojo" digital twin, correcting the camera/LiDAR calibration over three iterations; the first approach was off by about 0.5 meters and 0.4 meters too high.
A team built a proof-of-concept prototype for automated packing in a dark store using the open-source LeRobot SO-101 (SO-ARM101) 3D-printed robotic arm, winning first place at a hackathon. Targeting the high labor costs and slow picking of instant retail, the project used the Hugging Face-supported LeRobot ecosystem to complete teleoperation, data collection, and imitation learning within the hackathon timeframe, enabling the arm to reliably perform pick-and-place tasks, and released a 1-minute demo video.
ANet Drone4D V0.2 builds on the V0.1 multi-agent 4D simulation platform with an interactive sandbox, quantified target perception, nest orchestration and 7×24 zero-gap relay surveillance; a public demo site is now live. The release offers six classes of quantified perception targets, seven task types, a REST API and a Python SDK; 44 of 55 items passed in the second round of acceptance testing. Known issues include a limit of 2 concurrent sandboxes on the public site, short surveillance gaps for daytime random-walk targets, and a median frame time of 50 ms without a hardware GPU.