EN
Quadrotors learn to cross moving gaps in simulation
Learning Visuomotor Quadrotor Flight Through Dynamic Narrow Gaps via Differentiable Simulation · arXiv:2610.37557
In plain words
A quadrotor crossing a moving opening has to point itself through the gap and choose when to cross. The paper introduces a way to train its controller by running simulated flights and using the flight path to show how to adjust its movements. It does not need a planner or demonstration flights.
In the training simulator, the visual controller crossed successfully in 99.3% of 1,000 moving-gap trials. That result is for simulated trials, not real flights.?
The research
A moving gap makes flight depend on both where the quadrotor is and when it reaches the opening. Can a controller learn that precise behavior from onboard observations, without a planner or demonstration flights?
The method trains quadrotor controllers through differentiable simulated flights, guiding them to align with and cross moving gaps.
The authors train quadrotor controllers by differentiating through simulated flights, so position, alignment, and collision costs can guide each update. For moving gaps, this guidance follows the aperture as it changes pose.
The approach supports controllers using either privileged simulator state or onboard visual and inertial features, without a planner, teacher, or demonstrations.
In the training simulator, the visual controller succeeded on 99.3% of 1,000 moving-gap rollouts. In a separate state-based moving-gap comparison, the reported means show PPO used 19 times as many transitions and 45.5 times as much training time as BPTT.
PPO's mean covers four converged seeds; a fifth was censored at 30,000 epochs. PPO used tuned weights and a reward curriculum, while BPTT used fixed weights, so this result compares full training pipelines rather than gradient estimators alone.
The results cover simulated static gaps and planar pendulum-driven motion, plus ten verifiable real-world flights. The physical trial count is small, and the tested visual setup used LED-marked gap corners; the results do not establish performance for other kinds of openings or motion.
What the evidence shows
- A crossing needs timing
The paper introduces a method for training quadrotor controllers to cross narrow gaps, including gaps that move. The central challenge is that the vehicle must both fit through the opening and time its crossing, while onboard observations give only partial information near the gap.
The authors ask whether it can learn this behavior from simulation without demonstrations, a planner, or a teacher.
- Guidance through simulation
The method differentiates through simulated flights so that training can use information about how a sequence of movements changes the outcome. Its objective gives guidance on progress, centering, vehicle alignment, and collision avoidance; for moving gaps, the relevant geometry changes with the opening.
The visual controller receives engineered features from projected gap corners, together with inertial measurements.
- A diagnostic comparison
A separate planar navigation experiment tested whether this learning signal helps in a narrow-passage task. With the same reward, network architecture, and optimizer settings, pathwise gradients solved the task using one rollout per update, while the score-function approach failed with up to 4,096.
This supports the method's motivation for that tested task, rather than a general claim about every control problem.
- Simulation performance
For moving gaps in the training simulator, the visual controller succeeded on 99.3% of 1,000 rollouts. In the separate state-based moving-gap comparison, the reported means show PPO used 19 times as many transitions and 45.5 times as much training time as BPTT.
PPO's mean covers four converged seeds; a fifth was censored at 30,000 epochs. PPO also used tuned weights and a reward curriculum, while BPTT used fixed weights, so this compares training pipelines rather than gradient estimators alone.
- Transfer has limits
The visual controller also crossed in a higher-fidelity simulator and in physical flights, but its success rate was lower in the higher-fidelity simulator than in the training simulator. Static-only controllers performed poorly on moving gaps in the tested transfer comparison.
These results support training for moving-gap traversal and some transfer beyond the training simulator, while leaving broader robustness untested.
Deep
- Original paper authors
- Riccardo Bianco, Yunfan Ren, Jiaxu Xing, Davide Scaramuzza
- Submitted to arXiv
- License
- arXiv non-exclusive