SUMMARYIn typical calibration methods (kinematic or non-kinematic) for serial industrial robot, though measurement instruments with high resolutions are adopted, measurement configurations are optimized, and redundant parameters are eliminated from identification model, calibration accuracy is still limited under measurement noise. This might be because huge gaps still exist among the singular values of typical identification Jacobians, thereby causing the identification models ill conditioned. This paper addresses such problem by using new identification models established in two steps. First, the typical models are divided into the submodels with truncated singular values. In this way, the unknown parameters corresponding to the abnormal singular values are removed, thereby reducing the condition numbers of the new submodels. However, these models might still be ill conditioned. Therefore, the second step is to further centralize the singular values of each submodel by using a matrix balance method. Afterward, all submodels are well conditioned and obtain much higher observability indices compared with those of typical models. Simulation results indicate that significant improvements in the stability of identification results and the identifiability of unknown parameters are acquired by using the new identification submodels. Experimental results indicate that the proposed calibration method increases the identification accuracy without incurring additional hardware setup costs to the typical calibration method.

中文翻译：

---

使用等奇异值识别模型稳定标定六自由度串行机器人

总结在典型的串行工业机器人标定方法（运动学或非运动学）中，虽然采用了高分辨率的测量仪器，优化了测量配置，消除了识别模型中的冗余参数，但在测量噪声下，标定精度仍然有限​​。这可能是因为典型识别雅可比矩阵的奇异值之间仍然存在巨大差距，从而导致识别模型病态。本文通过使用分两步建立的新识别模型来解决这个问题。首先，将典型模型划分为具有截断奇异值的子模型。这样就去除了异常奇异值对应的未知参数，从而减少了新子模型的条件数。然而，这些模型可能仍然是病态的。因此，第二步是通过使用矩阵平衡的方法进一步集中每个子模型的奇异值。之后，所有子模型都条件良好，与典型模型相比，获得了更高的可观测性指标。仿真结果表明，采用新的识别子模型后，识别结果的稳定性和未知参数的可识别性显着提高。实验结果表明，所提出的校准方法提高了识别精度，而不会对典型校准方法产生额外的硬件设置成本。与典型模型相比，所有子模型都具有良好的条件并获得了更高的可观测性指标。仿真结果表明，采用新的识别子模型后，识别结果的稳定性和未知参数的可识别性显着提高。实验结果表明，所提出的校准方法提高了识别精度，而不会对典型校准方法产生额外的硬件设置成本。与典型模型相比，所有子模型都具有良好的条件并获得了更高的可观测性指标。仿真结果表明，采用新的识别子模型后，识别结果的稳定性和未知参数的可识别性显着提高。实验结果表明，所提出的校准方法提高了识别精度，而不会对典型校准方法产生额外的硬件设置成本。