RoboSignal Daily · 2026-10-28 · Wednesday
ROBOSIGNAL
RynnVLA-Lat, Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions
This research presents RynnVLA-Latent, a framework that enables scalable pretraining of robotic manipulation policies by learning 3D-grounded latent actions from heterogeneous interaction videos. The method avoids reliance on native robot controls and leverages visual transitions, future-feature prediction, and camera-motion decoupling to capture control-relevant motion. The results show strong performance on multiple benchmarks, including LIBERO, LIBERO-Plus, VLABench, and RoboTwin 2.0, with success rates exceeding those of competing methods in several real-world tasks.
Research
arXiv Robotics — research abstractsRynnVLA-Lat, Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions
This research presents RynnVLA-Latent, a framework that enables scalable pretraining of robotic manipulation policies by learning 3D-grounded latent actions from heterogeneous interaction videos. The method avoids reliance on native robot controls and leverages visual transitions, future-feature prediction, and camera-motion decoupling to capture control-relevant motion. The results show strong performance on multiple benchmarks, including LIBERO, LIBERO-Plus, VLABench, and RoboTwin 2.0, with success rates exceeding those of competing methods in several real-world tasks.
AGIBOT Launches Genie Studio 2.0 to Bring Robots into Real Workplaces
AGIBOT has launched Genie Studio 2.0, an integrated platform for developing, deploying, and operating embodied AI applications. The platform supports tasks such as precision assembly and fleet management in real-world environments, with features including model development, task orchestration, and fleet operations. It has been validated in manufacturing and rail transit projects, achieving high success rates and operational efficiency.
Tech Xplore — RoboticsArtificial skin helps robots grasp pencil lead and pull away from needles
Researchers have developed an artificial skin for robots that allows them to grasp fragile pencil lead and pull away from needles. The system uses a neural-hierarchical processing architecture to manage high-density sensing data efficiently, achieving 98.2% accuracy in object classification after data compression. The work demonstrates the robot's ability to grasp fragile objects like pencil lead and avoid harmful contact with needles.
NVIDIA Releases ONNX Inference Exports for Metatron G1 Scene/Recovery Policy
NVIDIA has released ONNX inference exports of the Metatron G1 scene/recovery policy, which are used in the Metatron browser demo and G.1 Racer. These exports are not interchangeable and require specific runtime interfaces. The models are intended for interactive humanoid control and evaluation, with deployment options beyond the browser, but they are not validated for real-world hardware use.
arXiv Robotics — research abstractsCrazyflow: An Accurate, GPU-Accelerated, Differentiable Drone Simulator in JAX
Crazyflow is a new simulator for aerial robotics that combines high-fidelity dynamics with JAX for speed and differentiability. It supports both gradient-based and sampling-based methods, and enables in-flight reinforcement learning by training a recovery policy in under a second without domain randomization.
Humanoid Robot Fully Open-Source, From Mechanical Structure to Motion Algorithms Fully Open
RoboParty officially released the fully open-source humanoid robot RP1, covering mechanical drawings, electrical systems, self-developed joints, low-level control, simulation training, and deployment tools. RP1 has 24-30 degrees of freedom, supports multiple application scenarios, and provides resources such as URDF and MJCF for developers' research. The platform aims to provide an open development environment for research institutions, universities, and embodied model companies, promoting in-depth exploration and ecosystem construction of humanoid robot technology.
Robotics Data Collection: Why Physical Interactions Are Hard to Replicate
The article explores the challenges of training humanoid robots by highlighting the lack of available data and the difficulty of replicating physical interactions across different robot bodies and environments. It discusses the shift from data abundance to data scarcity, the rise of data factories, and the need for new data infrastructure and standards to support embodied AI development.
Data & Training
Beijing Humanoid Robot Innovation Center Releases RoboMIND Dataset with Over 50 Million Global Downloads
The RoboMIND dataset, released by the Beijing Humanoid Robot Innovation Center, has surpassed 50 million global downloads, becoming one of the most downloaded open-source datasets in embodied intelligence. It supports dual-arm coordination, mobile operation, dexterous hands, and high-fidelity tactile perception, with over 400,000 dual-arm operation trajectories and more than 700 real-world tasks. The dataset is part of a larger data base with 30+ scenarios and 180+ robots, supporting full data lifecycle from collection to deployment.
Beijing Humanoid Robot Innovation Center Open Sources RoboMIND Dataset with Over 50 Million Global Downloads
The Beijing Humanoid Robot Innovation Center announced that its open-source dataset RoboMIND has exceeded 50 million global downloads. Since its release in December 2025, the dataset has become one of the most downloaded in embodied intelligence, with download speeds increasing significantly over time. RoboMIND is the first dataset to support dual-arm coordination, mobile operation, dexterous hands, and high-fidelity tactile perception, and includes over 400,000 dual-arm operation trajectories covering more than 700 real-world tasks.
Robots & Hardware
Robotics 24/7 — Industry reportingLuxonis Launches OAK 4 FFC Board-Level Vision Computer for Humanoids, Drones, and Other Machines
Luxonis has launched the OAK 4 FFC, a modular vision computer that integrates five camera modules, compute, and AI processing on a single board. It supports customization for different robot applications, including humanoids and drones, with built-in calibration and synchronization tools.
Deployments
Sprit G2 Max Heavy-Lift Robot Achieves High-Precision Handling and All-Weather Operation
Sprit G2 Max is an industrial-grade heavy-lift robot capable of lifting 40 kg with a repeat positioning accuracy of ±0.5 mm, achieving 100% material recognition rate. The robot has been deployed on multiple production lines, marking the transition of physical AI from laboratory settings to real-world manufacturing environments.
Zh iyuan Releases 'Robot-Making' Robot, Steady Pace Without Sleep, Assembles One 'Companion' in 2.5 Minutes
Zh iyuan has launched the G2 Max, an industrial-grade heavy-duty robot with safety compliance, aimed at applications such as palletizing and material handling. It supports 7×24 continuous operation, with dual arms capable of carrying 40kg and a repeat positioning accuracy of ≤±0.5mm. The robot is compatible with various industrial materials and does not require extensive reprogramming. Its spatial adaptability has been applied in e-commerce warehousing, supporting mixed-line operations for multiple product types. It also performs the 'robot-making-robot' task on the Qiyuan T1 production line, maintaining a stable 2.5-minute assembly rhythm.
Briefs
- SceneAgent: An Agentic Pipeline from 3D and 4D Captures to Simulation EnvironmentsarXiv Robotics — research abstracts
- A Robot That Can Build Robots Goes Live. Zhuyuan Launches Heavy-Duty Embodied Robot精灵G2 MaxChinese robotics — Deployment discovery
- FlowHOI: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot ManipulationarXiv Robotics — research abstracts
- Zhou Chang Stays. SeedAI Core Figures Remain at the Center, Embodied Intelligence Startup Rumors Ignite 2026 AI CircleChinese robotics — Robot learning research
- Pretraining Knows What, Simulation Teaches How: Compositional Generalization in a Sim-Trained Robot AgentarXiv Robotics — research abstracts
- Zhiyuan Releases Heavy-Load Embodied Robot精灵G2 Max, Achieving 7×24 Hour Stable OperationChinese robotics — Deployment discovery
- Policy Shifts Focus to Specialized Robotics: The Rise of 'Four Small Tigers'RoboSpeak — WeChat
- 2026 Hardcore Achievements | Industry's First. Humanoid Robots Start Regular Duty, Achieve 98% Material Loading Success RateChinese robotics — Hardware and sensing
- UniManip: Open-World Robotic Manipulation via Agentic Operational GraphsarXiv Robotics — research abstracts
- Embodied Intelligence Industry Map: From Multi-Scene Repurchase to Single-Point Validation, Four Layers of Deep Differentiation in ImplementationChinese robotics — Robot learning research