KU Leuven's DUTIR Lets Robots Learn Motions Regardless of Sensor Placement
A new closed-form decomposition turns rigid-body motion and force trajectories into a representation that stays identical no matter which coordinate frame you pick.
ProTakeaways
- KU Leuven researchers introduce DUTIR, a coordinate-invariant representation for motion and force trajectories.
- The SU-decomposition splits a screw trajectory matrix into a frame-transformation part and an invariant part, in closed form.
- Works on both rigid-body twists and interaction-force wrenches, unifying motion and force analysis under one representation.
- A single hyperparameter L regularizes singularities like pure translations that broke earlier invariant methods.
- Prior experiments report 96.5% recognition on rigid-body motion datasets and 92.3% F1 for real-time gesture recognition.
- Reference implementation available at github.com/arnoverduyn/SU_decomposition.
Anyone who has trained a model on robot trajectories knows the frustration: record the same motion twice with slightly different sensor placements, and your algorithm treats them as two different things. A team from KU Leuven has published a mathematical fix, called the Dual-Upper-Triangular Invariant Representation (DUTIR), which encodes motion and force trajectories in a way that is completely independent of the coordinate system used to measure them.
The representation is computed in closed form, runs fast enough for real-time use, and has already beaten prior invariant methods on gesture and motion recognition tasks. The implementation lives on GitHub.
Why coordinate frames wreck your data
A trajectory in robotics is usually stored as a stream of six-dimensional vectors called twists (three angular plus three linear velocity components) or wrenches (three force plus three torque components). Both are examples of a more general object called a screw: a 6D vector attached to a physical axis in space.
The catch is that the numbers inside a twist or wrench change whenever you move your reference frame. Getting consistent results across different choices of coordinate system is the core challenge. Two common workarounds exist, and both are painful:
- Run a careful calibration procedure before every data collection, which is labour-intensive.
- Ignore the problem and hope your model generalizes, which usually fails on real datasets where the coordinate systems are misaligned and the relationships between them are unknown.
A cleaner path is to transform trajectories into a form that mathematically cannot change under a frame swap. Prior work has done this using local differential-geometric trajectory features that are inherently coordinate-invariant, but those methods blow up near singularities (like pure translations) and can be numerically fragile.