DiffDreamer: Coupling World Models with Differentiable Simulation for Agile Quadrotor Control

概要

Tasks such as quadrotor dynamic-gate racing require agents to anticipate external motion while maintaining precise high-frequency control. World-model-based reinforcement learning provides a natural solution by learning latent dynamics for future imagination, yet it often entangles environmental evolution and the robot's own dynamics in a single learned rollout, where model bias can affect policy optimization efficiency. We propose DiffDreamer, a differentiable reinforcement learning framework that tightly couples a learned world model with differentiable simulation for actor optimization. In one optimization step, the learned world model propagates latent states that capture external dynamics and history-dependent features, while the differentiable simulator propagates the robot's explicit physical state. We further decompose the actor objective into a critic-driven pathway for non-differentiable task rewards and a simulator-driven pathway for differentiable motion rewards. Experiments on aggressive tracking and dynamic-gate racing show that DiffDreamer improves training convergence over the model-free, model-based, and differentiable-RL baselines, and final performance over these baselines and classical planning and control methods. Additional experiments assess reward decomposition and the odometry-conditioned relative-state representation through ablations. We also provide proof-of-concept real-world flight experiments for validation.

参考リンク

• Paper ↗

リンク先のライセンス・再現性・データ条件をご確認ください。自動分類や実装の目安は保証ではありません。

実装を始めるプロンプト

参考リンクを確認したうえで、次の内容をコピーして利用できます。

# 論文・記事
DiffDreamer: Coupling World Models with Differentiable Simulation for Agile Quadrotor Control

# 概要
Tasks such as quadrotor dynamic-gate racing require agents to anticipate external motion while maintaining precise high-frequency control. World-model-based reinforcement learning provides a natural…

# 原典
https://arxiv.org/abs/2610.31617

上記の手法を実装してください。実行前に依存関係とライセンスを確認してください。