Event cameras offer high temporal resolution and low latency, making them ideal sensors for high-speed robotic applications where conventional cameras suffer from motion blur. However, their widespread adoption in robot learning is severely bottlenecked by the computational cost of simulating high-frequency event data during online training. In this work, we present Approximate Imitation Learning, a novel framework that fundamentally resolves this bottleneck, reducing policy training time for complex, agile drone flight from 52.44 hours to just 1.86 hours - a 28x computational speedup. Our key insight is to separate representation learning from policy search. We first leverage a large-scale offline dataset to learn a task-specific representation space. Subsequently, the policy is fine-tuned through online interactions that rely solely on lightweight state information, completely eliminating the need to render events during the active policy search phase. This training paradigm drastically reduces development overhead and enables event-based control policies to scale to complex environments. Furthermore, our approach eliminates the reliance on standard cameras or intermediate representations during deployment, mapping events directly to control commands. In simulation, our method matches or exceeds the performance of standard imitation learning baselines that require full online event rendering. Finally, we successfully validate the framework in the real world, demonstrating that a policy trained via this ultra-efficient paradigm enables a quadrotor to fly through highly cluttered environments at remarkable speeds of up to 9.8 m/s.
Figures & tables
Figure 2: During offline training, teacher actions aT supervise the event encoder FS and the shared action decoder A . The approximate student, using teacher observations, is trained by aligning its features h^S and actions a^S with the features hS and actions aS of the event student. During online training, teacher observations obtained from lightweight simulations are used to fine-tune the shared action decoder A . Consistency losses in both the action and feature spaces are applied during an additional offline alignment step. For clarity, the auxiliary decoder Do is omitted.
Figure 2Figure 3
Poisson Delta: 0.04
Poisson Delta: 0.05
Methods
Training Input
Success Rate
Mean Velocity [ sm ]
Success Rate
Mean Velocity [ sm ]
BC
Off. Event
0.80
8.53
0.70
8.13
DAgger [ 42 ]
On. Event
0.30
6.66
0.10
5.72
BC+DAgger
Off. Event + On. Event
1.00
9.00
0.90
8.47
Approx IL (Ours)
Off. Event + On. State
1.00
8.76
0.90
8.53
State Teacher (Ours)
On. State + Dist. Map
1.00
8.77
0.90
8.60
Table 1: Simulation results for methods using offline (Off.) and/or online (On.) training input.
Table 5
Figure 7: Simulation results with dynamic obstacles moving randomly at 0.5ms−11.5ms−1 .
Figure 8: The simulation environment consists of tall cylindrical obstacles placed in a natural scene (a). The scene is captured by a simulated event camera mounted on the quadrotor (b). For reference, the corresponding grayscale images are also shown (c).
Figure 9: The event student receives events in the form of event representations, which are first encoded into features hSti by the event student encoder FS . The student encoder FS consists of a recurrent EfficientNet [ 53 ] followed by the Projection Layers . These event features are then concatenated with the outputs of the Vector Layers , which encode auxiliary inputs, i.e., the direction command vˉti , the previous action aSti−1 , and, optionally, the state information sti . The combined features are fused in the Fusion Layer , and the final actions aSti are produced by the Action Head .
Figure 10: Our vectorized event generation first quantizes log intensities into bands and computes band differences between adjacent timesteps. Neighboring non-zero values are then differenced again. Events are triggered where this second subtraction yields zero.
Figure 11: The maximum and mean GPU memory usage and the required computation time are shown for our vectorized event generation (Vectorized) and ESIM’s GPU-based event generation (ESIM) across different numbers of environments. Our method leads to a 34% reduction in mean runtime (top) and significantly lowers peak GPU memory usage (bottom).
Figure 12: Our proposed method yields a policy that consistently flies longer trajectories in simulation than BC, the closest baseline applicable in the same training setting.
Table 12
Figure 13: The ground truth distance maps and corresponding predictions are shown for the M3ED dataset (A), simulation data (B), and the real-world test environment (C). Importantly, the distance maps and associated auxiliary decoders are used only during pretraining and are not involved in action inference of the event student.
Figure 14: The platform used in the real-world experiments features an NVIDIA Jetson TX2 running the event-based student, which processes events from a DVXplorer Micro sensor.
National Key Laboratory of Science and Technology on Multi-spectral Information Processing, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China