Why shared tasks demand shared understanding

When a human and a robot lift or position an object together, success depends on more than strength or precision. It relies on timing, anticipation and subtle cues humans exchange effortlessly. Humans instinctively match grip, pace and direction, adjusting continuously without speaking. Robots, however, struggle with this natural fluidity. Although many models can predict human motion, turning these predictions into responsive real-time control remains a major challenge.

Researchers at the University of Siegen set out to tackle a central question in collaborative assembly: can a robot control its motions directly through a trajectory-based latent space that reflects how humans naturally reach for and manipulate objects?

Rethinking how robots interpret human motion

Most latent space (LS) control approaches rely on frame-based models that compress high-dimensional motion into static snapshots. While useful for analysis, they often lose the fine detail needed for smooth joint manipulation. To overcome this, the team explored whether positional control is feasible in a latent space built from entire reach trajectories rather than single frames.

Using functional principal component analysis and probabilistic motion models, they mapped thousands of human reach motions into a compact latent representation. By analysing how these distributions shift when the intended target position changes, they assessed whether the latent space behaves predictably enough to support control.

What the study reveals

The results show that when target positions move linearly, the corresponding latent modes also shift along clear vectors. In many cases the distributions remain Gaussian, while more complex tasks produce Gaussian mixtures that reflect different natural motion styles. Crucially, predicted hand trajectories reached their intended targets with only small positional errors, suggesting that trajectory-based latent spaces can support positional control in collaborative tasks.

Towards more intuitive collaboration

These findings indicate that robots could one day react to human movements through a latent space that captures the natural variability of human behaviour. This would enable smoother, safer and more intuitive human-robot collaboration, bringing robotic co-workers a step closer to genuine teamwork.

Based on “Towards trajectory-based latent space control for human-robot collaboration”, published in Procedia CIRP (Vol. 134, 2025).