See how gain settings affect the sim-to-real gap in reaching tasks.
We study reaching tasks with a Franka Research 3 robot to directly isolate the motor-level sim-to-real gap. For each gain setting, we perform gain-specific system identification, train RL policies in the calibrated simulation, and deploy them on the real robot without fine-tuning.
Stiff, overdamped gains are easiest to model. The MSE after system identification is over an order of magnitude lower for stiff, overdamped gains. Tap any cell to see how the identified simulator tracks the real robot.
…but hardest to transfer. Paradoxically, stiff and overdamped gains produce the worst sim-to-real transfer. The dominant failure mode is high-frequency oscillation that amplifies small prediction errors into out-of-distribution states.
Zero-shot sim-to-real transfer. With compliant gains (left), the arm smoothly approaches the target. With stiff, overdamped gains (right), the same policy architecture produces violent high-frequency jittering — tap the sound icon to hear it.
Compliant × Underdamped
Stiff × Overdamped
Includes detailed system ID curves, trajectory overlays, and synchronized video comparison