Skip to main content
Back

Paderborn's FARS Teaches Drones to Race Using Plain-Language Fuzzy Rules

A new paper proposes fuzzy-logic reward shaping for reinforcement learning, cutting training epochs and variance on hard drone racing tracks.

Paper, Full Access
  • FARS uses fuzzy logic rules to shape RL rewards based on velocity and distance to target.
  • On Hard drone racing tracks, converges in ~100 epochs vs ~150 for potential-field baseline.
  • Mamdani variant reaches >90% gate success with the lowest variance across seeds.
  • Up to ~5% success rate improvement over non-fuzzy reward formulations on hard tasks.
  • Built on PPO in Isaac Lab with 2048 parallel environments and zigzag gate courses.
  • Read the full paper on arXiv; simulation-only, no sim-to-real yet.

Reward design remains one of the messiest parts of reinforcement learning. A team from Paderborn University's Automatic Control Group is pushing back with FARS, a fuzzy-logic reward shaping method that swaps hand-tuned weights for interpretable linguistic rules like if drone is far and fast, reward it. On harder drone racing benchmarks, the approach converges faster and with less seed-to-seed variance than a standard potential-field baseline.

The gains are modest but pointed. Success rates improve by up to roughly 5% over non-fuzzy reward formulations on the hardest tracks, and the Mamdani variant converges in about 100 training epochs instead of 150 for the baseline. Nothing here reshapes the field, but it makes a clean argument for putting a fuzzy inference layer between the environment and the RL reward.

Why crisp reward shaping breaks under pressure

RL performance in complex tasks stays highly sensitive to the reward function, since sparse or poorly balanced signals can hinder exploration and destabilize training. Conventional shaping offers denser feedback but leans on manually tuned combinations that are brittle and sensitive to task variations.

Drone racing stresses this nicely. The agent has to fly fast between gates that can be meters apart, then decelerate precisely to squeeze through a narrow opening. A single crisp distance-based shaping term forces you to pick constants that either encourage speed everywhere (and crash the drone into gates) or precision everywhere (and never finish a lap).