Reading Modes
Open Artifacts with links to supporting source material.
630 selected reportsLast 30 days 568 recordsAll coverage 2,235 records
NEURA Robotics participated in the 26th China International Industry Fair (CIIF) 2026, showcasing their robot family in Hall 8.1. The event featured applications in manufacturing and retail, including cobots performing welding and screwing tasks, and a humanoid robot handing out giveaways. The company also introduced Neuraverse and Neuron AI, emphasizing their platform for building and sharing robot applications.
Jiangsu Province has launched a three-year plan to build over 100 high-quality physical AI datasets by 2028, focusing on real-world data from industrial and manufacturing environments. The initiative emphasizes data sovereignty, open access, and standardized data collection, aiming to support the development of world models and embodied AI.
At the 2026 China Industrial Robot Expo, robot companies demonstrated two distinct approaches to embodied intelligence: one focused on open hardware platforms and toolchains for external AI integration, and the other on self-developed vertical models and process intelligence. Both paths aim to enhance autonomy and adaptability in industrial settings, with varying emphasis on collaboration versus proprietary development.
Sinovation tested GPT-6-Astra's performance in multiple simulation tasks and found that the model significantly improved task success rates with increased interaction steps and more attempts allowed. The article further discusses the challenges of embodied intelligence in real-world environments, proposing solutions to reduce inference costs, improve evaluation standards, and promote practical deployment, while also open-sourcing the SenseNova-RoboRSI framework to advance community research.
A research team from SenseNova and collaborators conducted a comprehensive analysis of GPT-6-Astra's embodied intelligence capabilities, achieving high scores in simulation benchmarks but facing significant challenges in real-world robot execution. The study highlights the gap between simulated and real-world performance, emphasizing the need for improved physical interaction, task planning, and execution efficiency.
Speedtong Juchuang has completed the NVIDIA HSB ecosystem integration for its second-generation full-solid-state laser radar E2 and ultra-long-range laser radar EM4, becoming the first company to achieve this. This integration enables direct sensor data communication with NVIDIA's computing platforms, enhancing efficiency and reducing development costs for global developers. The company's automotive and robotics businesses are both seeing strong growth, with significant market expansion and new product launches.
Dream Labs, founded by former NVIDIA DreamZero team members, has released the embodied model Dream-1, which supports zero-shot task execution across different robots. The model demonstrates strong generalization capabilities in both simulated and real-world environments, achieving the top score in the RoboDojo-RC Tier 1 test. Dream-1 learns to predict visual scenes and actions through a world action model (WAM), allowing it to perform tasks on previously unseen dual-arm robots without additional adaptation.
OccSim is a data-driven autonomous driving simulation framework that eliminates the need for HD maps or pre-recorded logs. It generates over 3000 continuous frames and maps spanning more than 9 kilometers, with the potential to extend beyond 110K frames. The system uses a static occupancy world model and a Layout Generator to synthesize diverse simulation data, achieving up to 74% zero-shot performance on unseen data when scaled.
This paper presents RoboEdit, a system that transforms human manipulation videos into physically plausible robot videos with aligned 3D hand states. The method includes a pipeline for cross-embodiment adaptation and a large-scale dataset of 174K video pairs. Experiments show state-of-the-art editing quality and support for real-world robot control policies.
TRACT addresses the challenge of action chunking in robot imitation learning by introducing a method that separates current phase authority from future action routing. It uses a task-local graph to constrain phase authority and a cumulative boundary distribution to route future queries through phase-specific paths. The approach improves success rates and reduces phase ambiguity and stalls in contact-rich manipulation tasks.
This paper presents MIRA, an open-source micro-UAV designed for indoor research. Built around a 3D-printed PLA airframe with detachable arms and a containerized software package, MIRA supports modular customization and rapid development. The system is evaluated through autonomous flight tests, showing low-latency communication and stable structural behavior during dynamic maneuvers. All design files and software are released openly to support replication and extension by other laboratories.
This research presents MPAIL2, the first method to learn visual manipulation in the real world from scratch using observational learning via Inverse Reinforcement Learning from Observation (IRLfO). It achieves 82% performance in pick-and-place tasks within 40 minutes, surpassing reinforcement learning and behavior cloning approaches with comparable interaction and demonstration budgets.
This research presents Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. The model is trained on 10k hours of human egocentric data and 3,000 hours of robot data with unified multimodal annotations. It achieves state-of-the-art performance across five cross-embodiment benchmarks and supports real-time VR interaction after distillation.
This research introduces PhantomTac, a physical-layer security attack on magnetic tactile sensors used in robotic manipulation. By attaching a thin, battery-free magnetic tag to an object, an attacker can spoof force readings and compromise robotic control. The attack is 'deploy-and-forget' and can be concealed with a barcode sticker. Real-world evaluations on robotic arms show over 95% attack success rate under certain conditions, with a proposed defense mechanism based on multimodal consistency.
This research proposes VLA-Rolling, a framework that reuses action tails from vision-language-action models to enhance control efficiency. By adapting denoising budgets and refining actions through confidence-guided steps, the method achieves faster inference and improved success rates across multiple benchmarks.
DexOAK is a learning-based framework that generates contact-consistent robot hand references from human motion-capture data. It uses an interaction-graph Laplacian objective to reproduce object-relative interaction geometry, and a feed-forward network conditioned on human keypoints and object point clouds to learn retargeting. Object-aware finetuning refines the network under contact, non-penetration, and self-collision losses, improving performance on tasks like OakInk-v2 and ARCTIC. The model transfers to unseen data and enables dexterous robot data generation from human data.
This research proposes Self-Crystallizing Behavior Trees (SCBT), a method that integrates foundation models into deterministic control systems for autonomous drones. SCBT uses an LLM as a compiler to generate and refine control patches during runtime, which are then crystallized into deterministic behavior trees. The approach demonstrates improved mission success rates compared to unpatched and hand-coded trees, though it faces challenges in correctly attributing certain failure scenarios.
This research introduces T(R,O) Grasp++, a scalable diffusion-based framework for cross-embodiment dexter, grasping in cluttered environments. The model achieves an average success rate of 88% across eight dexterous embodiments, with 0.3s inference latency and improved sample efficiency.
This research proposes Skill-RSI, a framework for creating reusable manipulation skills for embodied agents through recursive self-improvement. Skills are compiled in simulation and run as standalone code or under a cheap caller, achieving high success rates on new layouts. The study evaluates performance across 7 RoboTwin 2.0 tasks, demonstrating the effectiveness of inheritance over from-scratch rewriting and the cost-efficiency of frozen skills.
This research presents LENS, an edge scene-memory acceleration system for resource-constrained robots and autonomous vehicles. LENS runs on an AMD KR260 FPGA/ARM platform, combining an INT16 SCR accelerator and a fixed-point 3DGS renderer to enable efficient localization and rendering without a GPU. The system achieves performance comparable to GPU-based methods while reducing power consumption by up to 5.7x.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).