Aly Magassouba

Rigid-body mechanics can be directly embedded into transformer layers, achieving state-of-the-art manipulation performance with 27x fewer parameters and unparalleled robustness to geometric perturbations.

Turns out, a single robot push, analyzed with a physics-informed Transformer, is enough to estimate an object's mass and friction, even without force sensors.

DRL can now outmaneuver classical planners for double-Ackermann-steering robots, thanks to a novel reward function design that overcomes non-holonomic constraints.