世界模型原生新一代范式！极佳视界斩获全球第一后，GigaBrain-0.5M*再进化

Core Insights - The article introduces the next-generation embodiment world model, GigaBrain-0.5M*, which follows the success of GigaBrain-0.1, achieving the top rank in RoboChallenge globally [1][3]. Group 1: Model Performance - GigaBrain-0.5M* has demonstrated zero-error performance over several hours in real robot tasks such as folding clothes, brewing coffee, and industrial box folding [3][4]. - In comparison with mainstream imitation learning and reinforcement learning baseline methods like AWR and RECAP, GigaBrain-0.5M* shows a nearly 30% improvement in task success rate, achieving close to 100% success in complex multi-stage tasks [10][11]. Group 2: Learning Mechanism - The model employs a world model-conditioned approach, enhancing robustness in long-term tasks by predicting future states and values as input conditions [4][8]. - GigaBrain-0.5M* incorporates a Human-in-the-Loop continuous learning mechanism, allowing for iterative training based on human-curated model trajectories, leading to a closed-loop learning process of "action - reflection - evolution" [4][8]. Group 3: Data Utilization - The foundational model, GigaBrain-0.5, was pre-trained on a total of 10,931 hours of diverse robot operation data, with 61% generated from the in-house GigaWorld high-fidelity synthetic model [17]. - The introduction of vast amounts of data significantly enhances the model's task coverage and strategy robustness, enabling better generalization in complex, long-duration tasks [17]. Group 4: Company Overview - The company, 极佳视界, focuses on embodiment intelligence and general robotics, providing integrated solutions for industrial, commercial, and household scenarios [21]. - The core team comprises top researchers from prestigious institutions and experienced executives from leading global companies, positioning the company as a pioneer in the world model technology space [21].