Most autonomous systems—from social robots to health trackers—operate on a hidden assumption: that the "human" they are interacting with moves like a healthy, able-bodied young adult. This assumption is baked into the datasets used to train these models, which are often populated by university students or professional actors.
However, human movement is not a single, standardized signal. It is a diverse spectrum influenced by age, physical ability, and social context. Recent research is shifting toward Inclusive Kinematics, a direction that seeks to bridge the "representation gap" by modeling the unique movement signatures of children, the elderly, and individuals with disabilities.
The Failure of the Standard Human Model
When a model is trained on a homogeneous demographic, it learns a narrow definition of "normal" movement. This leads to significant performance drops when the system encounters anyone outside that narrow slice.
For instance, François Modave and colleagues demonstrated in Mobile Device Accuracy for Step Counting Across Age Groups that popular fitness trackers consistently undercount steps in older populations. The reason is intuitive but often ignored in algorithm design: geriatric gait is often less vigorous and has different temporal signatures than the "standard" gait the devices were optimized for.
This isn't just a matter of counting steps; it's a matter of safety. In Text to Blind Motion, Hee Jae Kim and Eshed Ohn-Bar point out that models trained on sighted populations fail to generalize to the navigation strategies of blind individuals. A robot that expects a person to move based on visual cues will be dangerously unpredictable when interacting with someone using a white cane or a guide dog.
Modeling the Persona of Movement
To fix this, researchers are moving beyond simple action labels (like "walking" or "reaching") to capture the persona of the movement—the "how" rather than just the "what."
Mingyi Shi introduced MotionPersona: Characteristics-aware Locomotion Control, a framework that generates locomotion tailored to a character's age, physical build, and even emotional state. Instead of a generic walk cycle, the system can synthesize the specific gait of a "tired elderly man." This shift from generic to characteristic-aware modeling allows AI to better predict and simulate the movements of a diverse population.
Similarly, Kojiro Takeyama and Misha Sra developed LocoVR: Multiuser Indoor Locomotion Dataset in Virtual Reality to capture "proxemics"—the social rules of space. By using VR to collect data across diverse home environments, they provide the spatial data necessary for robots to navigate around people while respecting the different personal space requirements of children versus adults.
Bridging the Data Desert
The primary obstacle to inclusive kinematics has been a "data desert." Collecting high-fidelity motion data from vulnerable or specialized populations is logistically difficult and often raises privacy concerns.
New multimodal approaches are attempting to solve this by using privacy-preserving sensors. Hongkai Chen and the team behind CUHK-X use a combination of mmWave and thermal sensors to understand human activity in multi-generational households. This allows models to learn the motor patterns of children and the elderly in natural settings without the intrusive presence of cameras.
In the realm of assistive technology, Guanjie Huang and Li Liu address the gap in Lend a Hand. They highlight that Cued Speech recognition models often fail because they are trained on professional "cuers" rather than everyday users with hearing impairments. Their work emphasizes that for AI to be truly inclusive, it must be trained on the "non-standard" movements of the target demographic itself.
Why Intuitive Physics Matters
Finally, building intuition for diverse movement requires moving from 2D pixels to 3D physics. Nidhi Seethapathi argues in Movement science needs different pose tracking algorithms that current benchmarks fail because they don't account for physical forces.
A child's center of mass and limb proportions are different from an adult's; an elderly person's balance strategies are dictated by different muscular constraints. By incorporating intuitive physics—as seen in the FORCE dataset by Ilya Petrov—models can begin to understand why different people move the way they do, leading to systems that are not just more accurate, but fundamentally more inclusive.