E2E-Fly: An Integrated Training-to-Deployment System for End-to-End Quadrotor Autonomy

Fangyu Sun; Fanxing Li; Linzuo Zhang; Yu Hu; Renbiao Jin; Shuyu Wu; Wenxian Yu; Danping Zou

cs.ROApr 14, 2026

E2E-Fly: An Integrated Training-to-Deployment System for End-to-End Quadrotor Autonomy

Authors: Fangyu Sun, Fanxing Li, Linzuo Zhang, Yu Hu, Renbiao Jin, Shuyu Wu, Wenxian Yu, Danping Zou

Organizations: School of Automation and Perception, Shanghai Jiao Tong University, China

Abstract

Training and transferring learning-based policies for quadrotors from simulation to reality remains challenging due to inefficient visual rendering, physical modeling inaccuracies, unmodeled sensor discrepancies, and the absence of a unified platform integrating differentiable physics learning into end-to-end training. While recent work has demonstrated various end-to-end quadrotor control tasks, few systems provide a systematic, zero-shot transfer pipeline, hindering reproducibility and real-world deployment. To bridge this gap, we introduce E2E-Fly, an integrated framework featuring an agile quadrotor platform coupled with a full-stack training, validation, and deployment workflow. The training framework incorporates a high-performance simulator with support for differentiable physics learning and reinforcement learning, alongside structured reward design tailored to common quadrotor tasks. We further introduce a two-stage validation strategy using sim-to-sim transfer and hardware-in-the-loop testing, and deploy policies onto two physical quadrotor platforms via a dedicated low-level control interface and a comprehensive sim-to-real alignment methodology, encompassing system identification, domain randomization, latency compensation, and noise modeling. To the best of our knowledge, this is the first work to systematically unify differentiable physical learning with training, validation, and real-world deployment for quadrotors. Finally, we demonstrate the effectiveness of our framework for training six end-to-end control tasks and deploy them in the real world.

Figures & tables

	Type	Hardware	Render	Sensors	Algorithm	Validation	Deployment	Number of Tasks
Flightmare [ 12 ]	simulator	$\times$	Unity	D,R,S,I	RL	HIL	$\times$	$\times$
RotorS [ 13 ]	simulator	$\times$	OpenGL	D,R,I	$\times$	$\times$	$\times$	$\times$
CrazyS [ 14 ]	simulator	$\times$	OpenGL	D,R,I	$\times$	$\times$	$\times$	$\times$
CrazySim [ 15 ]	simulator	$\times$	OpenGL	D,R,S,I	$\times$	$\times$	$\times$	$\times$
FlightGoggles [ 16 ]	simulator	$\times$	Unity	D,R,S,I	$\times$	HIL	$\times$	$\times$
FastSim [ 17 ]	simulator	$\times$	Unity	D,R,S,I,L	RL	HIL	$\times$	$\times$

TABLE I: Key characters comparison of state-of-the-art approaches in this area.

Fig. 1: The overview architecture of E2E-Fly. In the training phase, the state-based and vision-based inputs are acquired from VisFly. During this process, the reward function is designed according to the reward function manual, while the accurate dynamics model supports training both via RL and differentiable simulation. The trained policy can be directly transferred to AirSim via an internal interface for cross-platform sim-to-sim test. In the real-world deployment phase, we provide two quadrotor hardware platforms that support real-time hardware-in-the-loop simulation and onboard flight test, respectively. The transfer from policy outputs to low-level control commands is achieved through our open-source betaflight-ctrl package.

Alg.	Task	Reward
PPO	Hovering	$r_{t}^{\text{prog}}+r_{t}^{\text{act}}+r_{t}^{\text{v}}+r_{t}^{\text{anglev}}+r_{t}^{\text{goal}}+r_{t}^{\text{crash}}$
	Landing	$r_{t}^{\text{prog}}+r_{t}^{\text{act}}+r_{t}^{\text{v}}+r_{t}^{\text{anglev}}+r_{t}^{\text{goal}}+r_{t}^{\text{crash}}$
	Tracking	$r_{t}^{\text{prog}}+r_{t}^{\text{act}}+r_{t}^{\text{ori}}+r_{t}^{\text{v}}+r_{t}^{\text{anglev}}+r_{t}^{\text{goal}}+r_{t}^{\text{crash}}$
	Racing	$r_{t}^{\text{prog}}+r_{t}^{\text{act}}+r_{t}^{\text{goal}}+r_{t}^{\text{crash}}$
	Obstacle avoidance	$r_{t}^{\text{prog}}+r_{t}^{\text{ori}}+r_{t}^{\text{v}}+r_{t}^{\text{anglev}}+r_{t}^{\text{avoid}}+r_{t}^{\text{goal}}+r_{t}^{\text{crash}}$
BPTT	Hovering	$r_{t}^{\text{prog}}+r_{t}^{\text{act}}+r_{t}^{\text{ori}}+r_{t}^{\text{v}}+r_{t}^{\text{anglev}}$

TABLE II: The reward design instruction for different types of tasks.

Task	Curriculum Design
High speed obstacle avoidance	C1: obstacle-free point navigation
	C2: sparse obstacle navigation
	C3: high-speed collision-free flight
Racing with random obstacles	C1: obstacle-free racing
	C2: sparse obstacle racing
	C3: dense obstacle racing

TABLE III: The examples of curriculum designs for complex tasks.

Fig. 2: The hardware-in-the-loop simulation in E2E-Fly. It consists of a real quadrotor flying in a motion capture system combined with a photorealistic simulation of complex 3D environments. Multiple sensors can be simulated with minimal delays while virtually flying in various simulated scenes. Such hardware-in-the-loop simulation offers a modular framework for prototyping robust vision-based algorithms safely, efficiently, and inexpensively.

Fig. 3: Proposed quadrotor platforms: VIS-R and VIS-H. VIS-R is used for onboard experiments with a Radax X4 and an Intel D435i RGB-D camera. VIS-H supports offboard and hardware-in-the-loop experiments via a wireless data transmitter and a wireless video transmitter.

Fig. 4: The digital prototype employed for moment of inertia test. The model is constructed by accurately assigning the material and mass properties for each component based on the physical quadrotor. Following this, the inertia tensor is computed within the software environment, enabling the extraction of the three principal moments of inertia directly from its diagonal.

Fig. 5: The LY-5KGF test stand for motor system identification. It‘s capable of test propellers of various sizes and featuring software support for remote monitoring and parameter tuning.

Parameter	Value
mass $m$ [kg]	$0.47$ for VIS-H, $0.75$ for VIS-R
inertia $\bm{J}$ [kg $\cdot$ m 2 ]	[ $1.25$ , $1.28$ , $2.03$ ] $\times 10^{-3}$ for VIS-H, [ $1.41$ , $1.53$ , $2.05$ ] $\times 10^{-3}$ for VIS-R
maximum thrust per motor $\bm{f}^{max}$ [N]	$5.12$
arm length $\bm{r}$ [m]	$0.075$
motor time constant $k^{motor}$ [s]	$0.035$
thrust map $[k_{f2},k_{f1},k_{f0}]$ [N $\cdot$ s 2 /rad 2 , N $\cdot$ s/rad, N]	[ $4.04\times 10^{-7}$ , $2.56\times 10^{-5}$ , $-2.62\times 10^{-2}$ ]

TABLE IV: Accurate parameters of our hardware platforms

Task Type	Target Task	Randomization Parameters
State-based	Hovering	$\bm{p},\bm{v}$
	Landing	$\bm{p}$
	Racing	$\bm{p}$
	Tracking	$\bm{p},\bm{v}$
Vision-based	Racing with obstacles	$\bm{p},\bm{p}_{\text{gate}},\bm{q}_{\text{gate}},\bm{p}_{\text{ob}}$
Vision-based	Visual landing	$\bm{p},\bm{p}_{\text{size}},\bm{p}_{\text{shape}}$

TABLE V: Domain randomization parameters for specific tasks.

Sensor Type	Classes
Vision	Gaussian noise
	Salt & pepper noise
	Poisson noise
	Speckle noise
	Redwood depth noise
Imu	Gaussian white noise

TABLE VI: The type of noise provided in E2E-Fly [ 21 , 28 ] .

Fig. 6: Examples of benchmark training scenarios constructed in the E2E-Fly. We present the simulation performance across various tasks, demonstrating that all policies trained via differentiable simulation and RL can accomplish their objectives.

Fig. 7: Training rewards comparison between PPO and BPTT. It is evident from the figure that BPTT exhibits faster convergence, higher sample efficiency, and achieves a higher reward compared with PPO.

	FPS [its/s]				time steps [its]				Time [s]
	Hovering	Landing	Tracking	Racing	Hovering	Landing	Tracking	Racing	Hovering	Landing	Tracking	Racing
PPO	$\approx 20000$	$\approx 20000$	$\approx 20000$	$\approx 15000$	5e7	1e8	5e7	1e8	$\approx 2500$	$\approx 5000$	$\approx 2500$	$\approx 6600$
BPTT	$\approx 13000$	$\approx 13000$	$\approx 24000$	$\approx 20000$	1e7	1e7	2e7	5e7	$\approx 750$	$\approx 750$	$\approx 800$	$\approx 2500$

TABLE VII: We record the training FPS, the total time steps required to converge, and the overall training time for PPO and BPTT on the four baseline tasks. The FPS is measured with 100 parallel environments.

	Parameter	Value
BPTT	optimizer	Adam
	learning rate	1e-3 decay to 1e-5
	discount factor	0.99
	horizon length	96
	replay buffer size	100000
	batch size	25600

TABLE VIII: Parameters for policy training.

Task	Observation	Reward	Example Parameters
Hovering	$(\bm{p}_{t},\bm{q}_{t},\bm{v}_{t},\bm{a}_{t})\in\mathbb{R}^{13}$	$r_{t}^{\text{prog2}}+r_{t}^{\text{ori2}}+r_{t}^{\text{v1}}+r_{t}^{\text{anglev1}}$ (D)	$\lambda_{2}=0.01,\lambda_{10}=0.0001,\lambda_{13}=0.002,\lambda_{17}=0.002$
Landing	$(\bm{p}_{t},\bm{q}_{t},\bm{v}_{t},\bm{a}_{t})\in\mathbb{R}^{13}$	$r_{t}^{\text{progxy}}+r_{t}^{\text{vz}}+r_{t}^{\text{anglev1}}$ (D)	$\lambda_{3}=0.04,\lambda_{16}=0.1,\lambda_{17}=0.001$
Tracking	$(\bm{p}_{t}^{i=1-10},\bm{q}_{t},\bm{v}_{t},\bm{a}_{t})\in\mathbb{R}^{40}$	$r_{t}^{\text{prog2}}+r_{t}^{\text{ori2}}+r_{t}^{\text{v1}}+r_{t}^{\text{anglev1}}$ (D)	$\lambda_{2}=0.02,\lambda_{10}=0.001,\lambda_{13}=0.002,\lambda_{17}=0.002$
Racing	$(\bm{p}_{t}^{1},\bm{p}_{t}^{2},\bm{q}_{t},\bm{v}_{t},\bm{a}_{t})\in\mathbb{R}^{16}$	$r_{t}^{\text{prog1}}+r_{t}^{\text{ori2}}+r_{t}^{\text{v1}}+r_{t}^{\text{anglev1}}$ (D)	$\lambda_{1}=0.9,\lambda_{10}=0.001,\lambda_{13}=0.002,\lambda_{17}=0.002$

TABLE IX: The observation space and reward function of state-based tasks.

Fig. 8: Success rate and task error under different reward settings training via PPO. The figure presents the ablation study evaluating the effects of the individual reward components trained with PPO. The

r^{\text{full}}

represents the configuration using all reward components listed in Table II , while w/o

r^{\text{v}}

,w/o

r^{\text{prog}}

,w/o

r^{\text{act}}

,w/o

r^{\text{anglev}}

,w/o

r^{\text{sparse}}

denote the cases without linear-velocity reward, progress reward, action-smoothness reward, angular-velocity reward, and sparse reward, respectively. For the racing task, only the

r^{\text{full}}

, w/o

r^{\text{prog}}

, w/o

r^{\text{sparse}}

, w/o

r^{\text{anglev}}

reward curves are included.

Fig. 9: Success rate and task error under different reward settings training via BPTT. The figure presents the ablation study evaluating the effects of the individual reward components trained with BPTT. The

r^{\text{full}}

represents the configuration using all reward components listed in Table II , while w/o

r^{\text{v}}

,w/o

r^{\text{prog}}

,w/o

r^{\text{anglev}}

denote the cases without linear-velocity reward, progress reward, action-smoothness reward, angular-velocity reward, and sparse reward, respectively. For the racing task, only the

r^{\text{full}}

and w/o

r^{\text{prog}}

reward curves are included.

Fig. 10: Simulation result of vision-based landing. Subfigures (a), (b), and (c) depict the performance of the identical policy while landing on triangular, circular, and square landing pads, respectively. The corresponding segmentation maps from the downward-facing camera are displayed on the right side of each subfigure.

Fig. 11: Simulation result of racing with obstacles. Subfigures (a), (b), and (c) present simulation results obtained in S-shaped, 3D circle, and J-shaped race tracks, respectively, with maximum velocities achieving over 10 m/s.

	Visual landing	Racing with obstacles
Observation	$(\bm{p}_{t},\bm{q}_{t},\bm{v}_{t},\bm{a}_{t})\in\mathbb{R}^{13}$ & ${segmentation}\in\mathbb{R}^{64\times 64}$	$(\bm{p}_{t}^{1},\bm{p}_{t}^{2},\bm{q}_{t},\bm{v}_{t},v_{d},\bm{a}_{t})\in\mathbb{R}^{17}$ & $depth\in\mathbb{R}^{64\times 64}$
Reward	$r_{t}^{\text{progxy}}+r_{t}^{\text{ori2}}+r_{t}^{\text{vz}}+r_{t}^{\text{anglev1}}$ (D)	$r_{t}^{\text{prog1}}+r_{t}^{\text{act3}}+r_{t}^{\text{ori1}}+r_{t}^{\text{avoid1}}+r_{t}^{\text{goal}}+r_{t}^{\text{crash}}$ (PD)
Parameter	$\lambda_{3}=0.04,\lambda_{10}=0.003,\lambda_{16}=0.1,\lambda_{17}=0.001$	$\lambda_{1}=0.9,\lambda_{7}=0.025,\lambda_{8}=0.002,\lambda_{9}=0.04,\lambda_{18}=0.01,\lambda_{21}=5,\lambda_{22}=4$

TABLE X: The observation space and reward function of vision-based tasks.

Fig. 12: Reward and success rate curves of visual tasks. From left to right, the figure depicts the reward and success rate curves for racing with obstacles and visual landing.

Fig. 13: The step-response alignment. We align the step-response of angular velocity and the mass-normalized thrust along the z-axis.

Fig. 14: Policy response after step-response alignment. The corresponding policy deployment without requiring additional alignment, yielding a response that closely matches the simulation.

Fig. 15: Zero-shot transfer from simulation to real world via our E2E-Fly system. Real-world performance demonstrates that the policy trained with our system can achieve zero-shot transfer from simulation to the real world.

Explore similar work

Mar 22, 2026

VisFly-Lab: Unified Differentiable Framework for First-Order Reinforcement Learning of Quadrotor Control

First-order reinforcement learning with differentiable simulation is promising for quadrotor control, but practical progress remains fragmented across task-specific settings. To support more systematic development and evaluation, we present a unified differentiable framework for multi-task quadrotor control. The framework is wrapped, extensible, and equipped with deployment-oriented dynamics, providing a common interface across four representative tasks: hovering, tracking, landing, and racing. We also present the suite of first-order learning algorithms, where we identify two practical bottlenecks of standard first-order training: limited state coverage caused by horizon initialization and gradient bias caused by partially non-differentiable rewards. To address these issues, we propose Amended Backpropagation Through Time (ABPT), which combines differentiable rollout optimization, a value-based auxiliary objective, and visited-state initialization to improve training robustness. Experimental results show that ABPT yields the clearest gains in tasks with partially non-differentiable rewards, while remaining competitive in fully differentiable settings. We further provide proof-of-concept real-world deployments showing initial transferability of policies learned in the proposed framework beyond simulation.

Fanxing Li, Fangyu Sun, Tianbao Zhang +5

Reinforcement Learning Quadrotor Control Multi-Task RL Differentiable Physics

Ask PDF

Apr 12, 2026

Simple but Stable, Fast and Safe: Achieve End-to-end Control by High-Fidelity Differentiable Simulation

Obstacle avoidance is a fundamental vision-based task essential for enabling quadrotors to perform advanced applications. When planning the trajectory, existing approaches both on optimization and learning typically regard quadrotor as a point-mass model, giving path or velocity commands then tracking the commands by outer-loop controller. However, at high speeds, planned trajectories sometimes become dynamically infeasible in actual flight, which beyond the capacity of controller. In this paper, we propose a novel end-to-end policy that directly maps depth images to low-level bodyrate commands by reinforcement learning via differentiable simulation. The high-fidelity simulation in training after parameter identification significantly reduces all the gaps between training, simulation and real world. Analytical process by differentiable simulation provides accurate gradient to ensure efficiently training the low-level policy without expert guidance. The policy employs a lightweight and the most simple inference pipeline that runs without explicit mapping, backbone networks, primitives, recurrent structures, or backend controllers, nor curriculum or privileged guidance. By inferring low-level command directly to the hardware controller, the method enables full flight envelope control and avoids the dynamic-infeasible issue.Experimental results demonstrate that the proposed approach achieves the highest success rate and the lowest jerk among state-of-the-art baselines across multiple benchmarks. The policy also exhibits strong generalization, successfully deploying zero-shot in unseen, outdoor environments while reaching speeds of up to 7.5m/s as well as stably flying in the super-dense forest. This work is released at https://github.com/Fanxing-LI/avoidance.

Fanxing Li, Shengyang Wang, Yuxiang Huang +5

Shanghai Jiao Tong University · Kunshan Duke University

Quadrotor Control Vision-Based Robot Control Robot Collision Avoidance RL for Robotics

Ask PDF

Feb 17, 2026

Efficient Knowledge Transfer for Jump-Starting Control Policy Learning of Multirotors through Physics-Aware Neural Architectures

Efficiently training control policies for robots is a major challenge that can greatly benefit from utilizing knowledge gained from training similar systems through cross-embodiment knowledge transfer. In this work, we focus on accelerating policy training using a library-based initialization scheme that enables effective knowledge transfer across multirotor configurations. By leveraging a physics-aware neural control architecture that combines a reinforcement learning-based controller and a supervised control allocation network, we enable the reuse of previously trained policies. To this end, we utilize a policy evaluation-based similarity measure that identifies suitable policies for initialization from a library. We demonstrate that this measure correlates with the reduction in environment interactions needed to reach target performance and is therefore suited for initialization. Extensive simulation and real-world experiments confirm that our control architecture achieves state-of-the-art control performance, and that our initialization scheme saves on average up to

73.5\%

of environment interactions (compared to training a policy from scratch) across diverse quadrotor and hexarotor designs, paving the way for efficient cross-embodiment transfer in reinforcement learning.

Welf Rehberg, Mihir Kulkarni, Philipp Weiss +1

Department of Engineering Cybernetics at the Norwegian University of Science and Technology, O.S. Bragstads Plass 2D, 7034, Trondheim, Norway

Robotic RL Cross-Embodiment Robot Learning Data-Efficient Robot Learning Reinforcement Learning Quadrotor Control

Ask PDF