As I was going through this paper (https://lnkd.in/gZ_jK5iA), my mind was trying to explore opportunities beyond sports analytics. Among the few possibilities was one futuristic but interesting one, which you can find at the end of this post. But first, let us try to understand what the authors are trying to do here:

The authors focus on predicting pitching speed in baseball using biomechanical data (kinematic features of joints during pitching). Why?

Because traditional models in biomechanics (e.g., Long Short‑Term Memory (LSTM) networks), though widely used, have limitations in interpretability (i.e., “black-box” behaviour) and in capturing complex spatial relationships between body joints.

Hence, the authors propose a hybrid model combining a Graph Neural Network (GNN) (to capture spatial relationships among joints) with a Gated Recurrent Unit (GRU) (to capture the temporal sequence of motion), essentially a “GNN-GRU”. They also embed interpretability via Layer‑Wise Relevance Propagation (LRP).

Overall, there are five key steps in their solution:

1. Feature extraction: For each pitch, kinematic angles for e.g., hip, knee, ankle, trunk, shoulders, were computed using visual3D or equivalent.

2. Graph construction: The human body is modelled as a graph: joints = nodes; bones/links = edges. The GNN processes this structure and learns spatial inter-joint dependencies.

3. Temporal modelling: The GRU processes sequences of these joint-node features across the pitching cycle (~101 time points per pitch) to capture dynamics.

4. Hyperparameter tuning: They used particle swarm optimization to choose model parameters.

5. Interpretability: They applied LRP to the trained GNN-GRU model to compute contribution rates of each feature (joint) at each time step in the pitch cycle, thus making the model’s decision mechanism more transparent.

One use case in the future can be in the area of humanoid bots’ safety and control. Once humanoid bots start penetrating homes in masses, you can extrapolate this approach to float a solution that monitors these robots for safety.

Essentially, by leveraging their posture and kinematics, they can alert if an action being taken by a bot can turn hazardous. These external Edge AI smart camera solutions can tether with internal Robotics AI to further fine-tune the kinematics of the bot. It can be leveraged in warehouses and manufacturing as well, basically wherever humanoid bots will be leveraged at scale.