ERA-42 - A large-scale end-to-end native robot model launched by Star Era

What is ERA-42?

ERA-42 is an end-to-end native robot model launched by Beijing Xingdong Era. Combined with its self-developed five-fingered dexterous hand, Xingdong XHAND1, it can complete more than 100 complex dexterity tasks. ERA-42 requires no pre-programming and possesses the ability to quickly learn new skills, mastering a new task within two hours using limited data. As the industry's first five-fingered dexterous hand embodied large-scale model, ERA-42 demonstrates powerful cross-modal capabilities, adaptability, and generalization ability, leading embodied large-scale models into the era of general dexterity manipulation and foreshadowing the vision of robots serving all industries and entering every household.

ERA-42's main functions

• DexterityWhen combined with the XHAND1 five-finger dexterity hand, it can perform more than 100 complex and dexterous tasks, such as using tools and grasping objects.
• Learn new skills quicklyIt can quickly learn and perform new tasks without pre-programming, and can learn new tasks with a small amount of data in less than 2 hours.
• Cross-modal capabilityThe model integrates multimodal information such as vision, language, touch, and body posture to achieve generalization ability for different tasks and environments.
• End-to-end executionThe entire process, from receiving full-modal data to generating the final output (such as decisions, actions, etc.), is completed based on a simple neural network link, without the need for human-designed features, pre-programming, or intervention processing steps.
• Understanding and predicting the physical worldAfter integrating the world model, it possesses the ability to understand the physical world and predict future action trajectories.

Technical principles of ERA-42

• Unified model generalizationBased on building a unified native model and integrating multiple modal information, it achieves the ability to generalize to different tasks and environments.
• End-to-end learningIt adopts an end-to-end learning approach, going directly from full-modal input to final output without the need for intermediate human intervention, thereby improving flexibility and development efficiency.
• Data-driven adaptation and generalizationBased on a large-scale video data learning strategy, the causal relationship is grasped using the results of the learned actions, thus achieving full generalization.
• World Model FusionThe goal is to integrate a world model into a native robot model, enabling it not only to move but also to understand the physical world.
• Joint learning of prediction and actionBy combining the denoising process, ERA-42 can learn how to improve predictions through actions, thereby enhancing the efficiency and accuracy of task execution.

Application scenarios of ERA-42

• Industrial AutomationUsed in automated production lines to perform complex assembly, testing, and maintenance tasks, thereby improving production efficiency and quality.
• Medical assistanceIn the medical field, it assists in delicate surgical procedures or in sample processing and experimental operations in the laboratory.
• Home servicesAs a household service robot, it completes household chores such as cleaning, cooking, and organizing, thereby improving the quality of life.
• Disaster relief: To carry out search and rescue missions at disaster sites, especially in environments that are difficult for humans to reach or that are dangerous.
• Logistics and distributionUsed in warehouse management and goods sorting to improve logistics efficiency and reduce labor costs.