Autonomous drone racing in complex environments requires agile, high-speed flight while maintaining reliable obstacle avoidance. Differentiable-physics-based policy learning has recently demonstrated high sample efficiency and remarkable performance across various tasks, including agile drone flight and quadruped locomotion. However, applying such methods to drone racing remains difficult, as key objective like gate traversal are inherently hard to express as smooth, differentiable losses. To address these challenges, we propose DiffRacing, a novel vector field-augmented differentiable policy learning framework. DiffRacing integrates differentiable losses and vector fields into the training process to provide continuous and stable gradient signals, balancing obstacle avoidance and high-speed gate traversal. In addition, a differentiable Delta Action Model compensates for dynamics mismatch, enabling efficient sim-to-real transfer without explicit system identification. Extensive simulation and real-world experiments demonstrate that DiffRacing achieves superior sample efficiency, faster convergence, and robust flight performance, thereby demonstrating that vector fields can augment traditional gradient-based policy learning with a task-specific geometric prior.
Figures & tables
Fig. 1: Compared to typical differentiable-dynamics-based methods that rely solely on dense differentiable loss functions to provide gradients, our framework integrates Attractive Vector Fields as a geometric prior for gate traversal, alongside standard loss functions to ensure safety.
Fig. 2: Overview of the DiffRacing Framework. The policy network ( π ) takes depth and state observations as input. The core components are: the differentiable simulator, which allows loss gradients to back-propagate to the network; the Attractive Vector Field module, which augments policy gradients during training. The detailed augmentation mechanism is provided in Sec. III-C ; and the Delta Action Model, which is trained to compensate for dynamics mismatch. Dashed lines indicate gradient flow. The orange lines represent the data flow during policy training, where the entire pipeline—including UAV dynamics rollout, image rendering, loss computation, and gradient backpropagation—is carried out fully within the simulator. The blue lines represent the data flow described in Sec. III-D , which includes data collection, Delta Action Model training, and subsequent policy fine-tuning.
Fig. 3: 3D Magnetic field visualization: (a) Top view of magnetic field; (b) Axonometric view of magnetic field.
Fig. 4: An Intuitive Illustration of Trajectory Self-Correction (Top View): From left to right, an overshooting trajectory is gradually refined under the continuous guidance of the Attractive Vector Fields.
Settings
λp
Success Cross
Success Rate
Max Velocity
W AVF
0.4
0.95
0.97
6.5
W/O AVF, W Lp
0.4
0.00
0.98
4.7
0.8
0.00
0.89
5.9
1.6
0.00
0.74
7.4
3.2
0.28
0.30
9.7
W/O AVF, W Lpnorm
1.6
0.00
0.95
5.0
TABLE I: Performance Comparison with and without AVF
Fig. 5: Performance comparison between different training schemes.
Method
Difficulty Level
Zigzag
Circular
Ellipse
SR
vmax
SR
vmax
SR
vmax
Ours w/ DA
0
10/10
7.1
10/10
6.4
10/10
7.1
3
10/10
7.0
10/10
6.2
10/10
6.4
6
10/10
6.1
10/10
5.8
9/10
6.3
9
7/10
6.1
9/10
5.6
7/10
5.9
Ours w/o DA
0
10/10
6.8
10/10
5.9
10/10
6.5
TABLE III: Sim-to-sim Performance
Fig. 6: Velocity components during real-world flight. The dashed line shows the velocity predicted by the simulation with Delta Action Model compensation, while the solid line represents the velocity measured by the motion capture system.
Fig. 7: Real-world experiments. The policy is deployed on a physical drone navigating (a-c) zigzag and (d) circular tracks with complex, obstacle-dense layouts that were unseen during training. The maximum velocities recorded in (a), (b), and (c) are 5.9, 6.4, and 6.0 m/s, respectively.
Autonomous drone racing has attracted increasing interest as a research topic for exploring the limits of agile flight. However, existing studies primarily focus on obstacle-free racetracks, while the perception and dynamic challenges introduced by obstacles remain underexplored, often resulting in low success rates and limited robustness in real-world flight. To this end, we propose a novel vision-based curriculum reinforcement learning framework for training a robust controller capable of addressing unseen obstacles in drone racing. We combine multi-stage cu rriculum learning, domain randomization, and a multi-scene updating strategy to address the conflicting challenges of obstacle avoidance and gate traversal. Our end-to-end control policy is implemented as a single network, allowing high-speed flight of quadrotors in environments with variable obstacles. Both hardware-in-the-loop and real-world experiments demonstrate that our method achieves faster lap times and higher success rates than existing approaches, effectively advancing drone racing in obstacle-rich environments. The video and code are available at: https://github.com/SJTU-ViSYS-team/CRL-Drone-Racing.
Fangyu Sun, Fanxing Li, Yu Hu +4
School of Automation and Perception, Shanghai Jiao Tong University, China · Faculty of Aerospace Engineering, Delft University of Technology, The Netherlands
Autonomous drone racing requires the tight coupling of perception, planning, and control under extreme agility. However, recent approaches typically rely on precomputed spatial reference trajectories or explicit 6-DoF gate pose estimation, rendering them brittle to spatial perturbations, unmodeled track changes, and sensor noise. Conversely, end-to-end learning policies frequently overfit to specific track layouts and struggle with zero-shot generalization. To address these fundamental limitations, we propose a fully onboard, vision guided optimal control framework that enables reference-free agile flight through arbitrarily placed and oriented gates. Central to our approach is Gate-SDF, a novel, implicitly learned neural signed distance field. Gate-SDF directly processes raw, noisy depth images to predict a continuous spatial field that provides both collision repulsion and active geometric guidance toward the valid traversal area. We seamlessly integrate this representation into a sampling-based Model Predictive Path Integral (MPPI) controller. By fully exploiting GPU parallelism, the framework evaluates these continuous spatial constraints across thousands of simulated trajectory rollouts simultaneously in real time. Furthermore, our formulation inherently maintains spatial consistency, ensuring robust navigation even under severe visual occlusion during aggressive maneuvers. Extensive simulations and real-world experiments demonstrate that the proposed system achieves high-speed agile flight and successfully navigates unseen tracks subject to severe unmodeled gate displacements and orientation perturbations. Videos are available at https://zhaofangguo.github.io/vision_guided_mppi/
Fangguo Zhao, Hanbing Zhang, Zhouheng Li +2
College of Control Science and Engineering, Zhejiang University, Hangzhou 310027, China
Autonomous drone racing is a fundamentally challenging regime for autonomous aerial robots, requiring time-optimal control while operating under persistent actuation saturation. While reinforcement learning (RL) has achieved human-level performance in this domain, current methods fail to generalize; policies trained on specific environments often crash immediately in unseen configurations. This failure reflects the intrinsic difficulty of zero-shot generalization in agile flight, arising from high-dimensional task variation and the tight coupling between safety and performance at high speeds. Existing approaches that improve generalization impose a substantial cost on flight speed: control policies must significantly degrade performance to achieve even modest levels of generalization. In this work, we propose a framework for zero-shot generalization in agile flight for RL-based drone racing. By combining task-aware switching based on learning progress with a physically informed procedural track generator, the framework produces a fast and robust generalist policy without test-time adaptation. Our method achieves strong zero-shot performance across a wide range of unseen racetracks in the real world, demonstrating a 7.4x improvement in generalization over the state-of-the-art approaches, while maintaining competitive racing speeds. We validate our method's results in both simulation and real-world settings, including a challenging vision-based, end-to-end control setting that operates without explicit state estimation, where all prior approaches fail to generalize.
Jonathan Green, Jiaxu Xing, Nico Messikommer +2
Robotics and Perception Group, University of Zurich