In a 26 October Scientific Reports study, a Crazyflie drone carried a 5-gram load on a 30-centimetre cable. Vicon tracking measured both.[1]
The team used Proximal Policy Optimisation, or PPO, in SimpleFlight to reward accurate payload position, uprightness and smooth control. The aircraft could move to serve the load’s path.[1]
As an illustration, picture carrying a bucket on a string: your hand follows one path while the bucket swings along another.[1]
A fixed policy can keep responding
A policy is a learned rule for choosing the next action from current observations. SimpleFlight’s 2025 methods paper describes a neural network mapping upcoming waypoints, speed and orientation to total thrust and turning-rate commands. A lower-level controller converts those commands into individual motor outputs.[2]
Training and flying are separate. In simulation, the algorithm repeatedly adjusts the rule in response to rewards. During flight, new measurements can produce new commands even when the rule itself stops changing. Fixed refers to the learned parameters, not a prerecorded flight.[2]
The SimpleFlight study also explains why rewarding smooth commands matters. A simulator may tolerate abrupt thrust changes that destabilise a physical aircraft. Penalising changes between successive commands helps address that mismatch, although excessive smoothing can restrict agility.[2]
Four physical tests
| Flight condition | Mean payload position error |
|---|---|
| Figure-eight at 0.5 m/s | 0.107 mSource 1 |
| Pentagram at 0.5 m/s | 0.136 mSource 1 |
| Figure-eight at 1 m/s | 0.173 mSource 1 |
| Figure-eight at 0.5 m/s with fan crosswind | 0.141 mSource 1 |
A separate hover test compared the two targets: mean error was 0.073 metres with payload-based training and 0.158 with drone-based training, five flights each.[1]
Sources & further reading
- Characterization of a fixed reinforcement learning policy for aerial robot with suspended payload under variable flight conditions Scientific Reports · 26 Oct 2026 · Accessed 28 Oct 2026
- What Matters in Learning a Zero-Shot Sim-to-Real RL Policy for Quadrotor Control? A Comprehensive Study IEEE Robotics and Automation Letters · 29 May 2025 · Accessed 28 Oct 2026