#USA #Pittsburgh #IROS 2026 #LumiBot #Shiyue Technology

LumiBot Challenges Data Bottleneck in Embodied AI at IROS 2026

At the IROS 2026 exhibition, LumiBot, a promising startup known in China as Shiyue Technology, made a significant impression with its cutting-edge advancements in robotic technology. Launched just over a year ago, LumiBot is gaining traction in the field of embodied AI, showcasing its latest products, including advanced finger-shaped visuo-tactile sensors and dexterous hand models tailored for complex manipulation tasks.

Addressing the Hardware Data Bottleneck

A critical aspect of LumiBot's presentation revolves around an enduring challenge in the field: the upper limit of data quality is stringently tied to hardware reliability. In the realm of embodied AI, obtaining high-quality real-world manipulation data remains an ongoing struggle. Currently, the industry faces a scarcity of public multimodal datasets, with samples featuring high-degree-of-freedom (DOF) hands equipped with tactile feedback amounting to merely tens of thousands. To effectively train scalable manipulation models, upward of tens of millions of hours of real-world operational data is needed. LumiBot aims to resolve this bottleneck by addressing the hardware aspect directly, utilizing their highly reliable, high-precision hands and visuo-tactile sensors to ensure that the data collected is standardized, clean, and ready for reuse from day one.

High-Reliability Hardware for Consistent Data Collection

LumiBot's innovative Lumi-Dex hand series supports prominent industrial communication protocols and harmoniously integrates with a variety of robotic arm platforms and simulation environments. This allows for a unified execution point for different embodied architectures. Among its offerings, the Lumi-Dex-21 is particularly noteworthy; it features an ingenious drive-control design that merges linkage with direct-drive mechanisms. Also in the product pipeline is LumiBot's flagship dexterous hand, boasting improved durability, longevity, and thermal management. With a self-developed active-cooling micro-joint module, this hand maintains operational temperatures around 50°C and can sustain over 5,000 hours of continuous no-load operation, making it capable of weeks-long uninterrupted data collection.

Precision Data Capture

In terms of data collection, the Lumi-Tac visuo-tactile sensor V2 enhances its predecessor with structural and optical improvements, allowing for greater coverage across various fingertip areas with enhanced durability against wear and deformation. Its specifications include an impressive spatial resolution of 10 microns, a force detection accuracy of 0.02 N, and a frame rate of 120 fps. Coupled with LumiBot's proprietary high-precision curved surface reconstruction algorithms, Lumi-Tac delivers comprehensive multimodal tactile data streams, essential for training robust models.

Streamlined Data Mapping and Model Integration

In its endeavor to streamline data acquisition, LumiBot has optimized its real-robot teleoperation collection pipeline. The company has achieved precise millimeter-level verification through its egocentric (Ego) data-to-robot mapping, ensuring consistent training input for its models. At the forefront is the Lumi-Mind manipulation foundation model, structured with a hierarchical Mixture of Transforms (MoT) architecture, effectively integrating vision, tactile feedback, language, and action modalities. A unique memory mechanism and tactile gating enable the model to engage in long-term task planning and make fine-tuned adjustments during intricate interactions.

Demonstrating Real-World Applications

During the IROS expo, LumiBot showcased the capabilities of its Lumi-Dex-11, a lower-DOF version of its dexterous hand, in executing complex tasks. The demonstration involved a sophisticated sequence of actions: picking up a stamp, executing precise stamping, retrieving a tray, distributing a card, and finally, returning both the tray and stamp. Notably, when human interference occurred, the model adeptly reassessed its task approach in real time, showcasing the combined prowess of LumiBot's hardware and software.

Future Directions for LumiBot

Founded in 2025, LumiBot has meticulously developed a comprehensive product line entirely in-house, from the motors and visuo-tactile sensors to the dexterous hands and manipulation models. This self-sufficient approach guarantees that the data collected aligns seamlessly with the requirements of model training, offering advantages that pure hardware or algorithmic companies cannot provide alone.

Moving forward, LumiBot aims to enhance the real-world applicability of its dexterous hands by transitioning standard data collection practices from research environments into actual industrial scenarios. The company is currently engaged in scenario validation and technical verification with industry leaders in automotive manufacturing, 3C electronics, and life sciences, systematically proving the scalability and reliability of its systems in real production settings.

In summary, LumiBot's advancements at IROS 2026 underscore its commitment to advancing robotic dexterous manipulation intelligence. With its focus on high-quality data acquisition and the reliability of its hardware, LumiBot is poised to reshape the landscape of embodied AI, empowering robots for efficient operation in dynamic and complex environments.