千寻智能高阳团队最新成果：纯视觉VLA方案从有限数据中学到强大的空间泛化能力

Core Viewpoint - The article discusses the introduction of a new strategy called "State-free Policy" in visuomotor control for robots, which enhances spatial generalization capabilities by removing state information from the input, relying solely on visual observations [1][10][24]. Group 1: State-free Policy Overview - The State-free Policy allows robots to perform tasks effectively even when the training data is strictly controlled, such as fixed desktop height and object positions [3][10]. - This policy is based on two key conditions: representing actions in a relative end-effector space and ensuring comprehensive visual input for task observation [11][13]. Group 2: Experimental Results - Extensive experiments demonstrated that State-free Policy significantly improves spatial generalization, achieving a success rate of 0.98 in height generalization and 0.58 in horizontal generalization for the pen insertion task [14][24]. - In challenging tasks like folding clothes and fetching bottles, State-free Policy outperformed state-based models, showcasing superior horizontal generalization capabilities [17][20]. Group 3: Advantages of State-free Policy - State-free Policy exhibits higher data utilization efficiency, maintaining performance even with limited data, unlike state-based strategies that tend to overfit [20][21]. - The policy also shows advantages in cross-platform adaptation, requiring less adjustment compared to state-based strategies, leading to faster convergence and higher success rates [21][22]. Group 4: Sensor Design Considerations - The research suggests reevaluating sensor designs, particularly the necessity of overhead cameras, as they may introduce performance issues due to changes in object positions [22][23]. - The findings indicate that using dual wide-angle wrist cameras can provide sufficient task observation without the overhead camera, maintaining high success rates in various scenarios [23].