HarmoAtt-IK: an adaptive multimodal feature fusion network for real-time neural inverse kinematics

Recent advances in neural networks have introduced a new paradigm for robotic inverse kinematics. However, existing methods remain limited by insufficient feature extraction and suboptimal integration of multi-source information, preventing them from achieving high accuracy, broad generalization, and real-time performance on robots with diverse and complex kinematic structures.

In this work, we propose HarmoAtt-IK, an adaptive multimodal neural inverse kinematics approach designed for real-time inference and zero-collection training. Built upon the CycleIK framework, the proposed method introduces a novel adaptive multimodal attention fusion mechanism (HarmoAtt) that dynamically integrates the complementary strengths of spatial, channel, and cross-dimensional attention. It employs a temperature-adaptive Softmax function coupled with a compact weight-generation network to perform multidimensional extraction and adaptive enhancement of input features. We further introduce a composite loss function integrating an improved Smooth-L1 loss, a sign-invariant quaternion loss, and a Shannon entropy regularizer to enhance training stability and overall accuracy. Leveraging forward differential kinematics, our method enables rapid, cross-platform deployment by generating training data solely from URDF models, eliminating the need for costly physical data collection and manual annotation.

Experimental evaluations on five humanoid platforms exhibiting substantial kinematic diversity demonstrate that HarmoAtt-IK attains maximum reductions of 76.4% in terminal positional error and 55.1% in rotational error relative to the baseline, while consistently improving the model's inference success rate across all tested platforms by up to 5.76 percentage points.

These results indicate that the proposed HarmoAtt-IK significantly outperforms baseline methods in both accuracy and reliability across diverse kinematic structures, highlighting the effectiveness of the adaptive multimodal attention mechanism and composite loss design. This further supports its potential for scalable, real-time deployment on a wide range of robotic platforms.