This work answers that with a rare yes.

The core problem has always been contact.
Soft contact models give smooth gradients but break realism.

Hard contact models are realistic but kill optimization due to discontinuities.

This paper proposes an analytically smooth contact formulation that keeps both.

• A quadruped locomotion policy trained entirely in a differentiable simulator
• Zero-shot transfer to real hardware
• More than 10× better sample efficiency compared to PPO

Differentiable simulation has promised better robot learning for years, but locomotion was the hard wall.

This is, to my knowledge, the first real sim-to-real legged locomotion result trained purely with analytic gradients.

If this scales, it could change how we train robots that rely on contact, not just legs, but manipulation too.

Learning Deployable Locomotion Control via Differentiable Simulation: