In recent years, imitation learning (IL) has established itself as a key tool in robotics and modeling of human gait. But a recent finding questions its reliability: When only motion data is used—with no information about external forces—estimated joint moments can be physiologically inconsistent. This paper explores the implications of this limitation and how the integration of kinetic data corrects for those deviations, opening the door to more robust applications in rehabilitation, sports analysis, and exoskeleton design.
The research that inspires this reflection compared two approaches: one based exclusively on motion (MOIL) and another that incorporates ground reaction forces (GRF) and pressure center (CoP) into the reward function (KAIL). The results showed that, although both achieve similar kinematic accuracy, MOIL produces considerable errors in GRF, CoP and joint moments with respect to the reference inverse dynamics. That is, a model trained only on trajectories can learn to move in a visually correct way, but without guaranteeing that the internal loads are realistic.
This discrepancy is not trivial. In contexts such as the assessment of pathological gait or the monitoring of intelligent prostheses, relying on kinematic estimates alone could lead to misinterpretations of the underlying biomechanics. For example, a patient might walk in an apparently normal pattern, but their joints might be enduring abnormal moments that accelerate joint wear and tear or generate harmful compensations. Data science applied to biomechanics therefore needs models that incorporate all relevant physical variables.
From a technical perspective, the solution is to enrich training sets with force and pressure cues, which poses capture and timing challenges. But the real added value is in the ability to process that information intelligently. This is where companies like Q2BSTUDIO bring their expertise in bespoke applications for research and development environments. Building platforms that integrate data from inertial sensors, force platforms, and motion capture systems requires bespoke software that ensures low latency and high fidelity in processing.
The trend towards artificial intelligence in the analysis of human movement is no longer a promise, but a reality. Deep learning models can learn latent representations that combine kinematics and kinetics, but they need a robust data architecture. At Q2BSTUDIO we develop AWS and Azure cloud services that scale the storage and compute of terabytes of time series, allowing you to train complex models without bottlenecks. In addition, our cybersecurity solutions ensure the protection of sensitive patient or athlete data, a prerequisite in clinical environments.
Another relevant aspect is the interpretability of the models. It is not enough to predict joint moments; Researchers need to understand which variables influence each prediction. Here, business intelligence services complement analytics: through interactive dashboards with Power BI it is possible to visualize correlations between kinematics, kinetics and performance metrics or injury risk. Q2BSTUDIO implements these custom dashboards, integrating data from sensors and AI models.
The aforementioned study also underlines the importance of AI agents in the simulation of imitation environments. An agent who learns to walk by reinforcement can benefit from a multi-criteria reward function that includes not only posture error, but also the plausibility of reaction forces. These types of architectures require careful development, where enterprise AI aligns with concrete biomechanical goals. At Q2BSTUDIO we collaborate with R+D teams to design training pipelines that integrate these signals efficiently.
Beyond research, these findings have commercial implications. Companies developing exoskeletons or virtual rehabilitation systems need to ensure that their algorithms generate safe movement patterns. If an exoskeleton is based solely on kinematics, it could apply joint pairs that do not correspond to the user's actual biomechanics, causing discomfort or even injury. Incorporating kinetic models reduces this risk and allows care to be personalised. Here it is key to have technology partners who understand both the physical and digital domains.
On the other hand, the application in the sports field is equally promising. Analyzing a runner's technique with video alone can lead to misleading conclusions; Adding reaction force data (measured with instrumented templates or platforms) gives a more complete picture of efficiency and injury risk. Real-time monitoring systems, powered by AWS and Azure cloud services, allow trainers and physiotherapists to access these indicators from anywhere.
To achieve all this, it is not enough to have advanced AI models; A foundation of bespoke applications is needed that connect sensors, databases, servers, and visualizations. Q2BSTUDIO designs and implements these integrated solutions, from data acquisition to delivering insights in interactive dashboards with Power BI. And all this with the cybersecurity guarantees required by health and sports standards.
In short, human movement data alone are insufficient to guarantee plausible biomechanics. The combination of kinematics and kinetics, along with a robust technology platform, is the path to more accurate models and safer applications. At Q2BSTUDIO we are committed to delivering those capabilities, helping research, health and sports organizations transform multidimensional data into trusted knowledge.
