Researchers at RISC Lab have developed a quadrotor interception policy that tracks and catches an intruder drone using only the direction to the target—no distance measurement required. By training the policy in a differentiable simulation, the approach achieves higher success rates and smoother flight than traditional reinforcement learning, opening the door to lightweight, camera-only pursuit drones.
What the Researchers Built
The team built an end-to-end neural network policy that directly outputs motor commands for a quadrotor interceptor. Unlike prior work that assumed the intruder’s full 3D position (including range) was known, this policy only receives the intruder’s direction vector—a unit vector pointing from the interceptor to the target—along with its own state (position, velocity, orientation). This makes the approach compatible with passive cameras and object detectors like YOLO, which provide bearing but not depth.
The policy is trained using Analytic Policy Gradients (APG) through a differentiable physics simulator of the quadrotor’s full 12-state dynamics. The training objective rewards proximity to the intruder at the end of the episode while keeping control effort low. The researchers compare their method against Proximal Policy Optimization (PPO) and a simplified point-mass dynamics model, demonstrating that full dynamics and differentiable simulation yield both faster learning and smoother interception trajectories.
Key Results
In simulation experiments, the APG-trained policy consistently outperformed PPO across multiple intruder speeds and trajectory patterns. The APG policy achieved higher interception success rates—defined as coming within a small radius of the intruder—and completed intercepts in fewer time steps. It also produced significantly smoother acceleration and jerk profiles, avoiding the abrupt, overshoot-prone maneuvers seen in PPO policies.
When trained on full quadrotor dynamics as opposed to a simplified point-mass dynamics (PMD) model, the resulting policy was more robust to the nonlinear effects of rotor thrust limits and drag. The analysis of acceleration and jerk revealed that the APG policy learned to modulate aggressiveness: it applies higher acceleration only when necessary to close the gap, then decelerates smoothly to avoid overshooting. This balance reduced the need for large corrective actions after a miss.
Although the paper reports no explicit percentage success rates in the provided text, the qualitative and comparative analysis makes a strong case for differentiable simulation as a superior training paradigm for this class of pursuit problems.
How It Works
The core innovation is the use of a differentiable quadrotor simulator. Traditional reinforcement learning treats the environment as a black box, sampling transitions to estimate gradients. APG instead backpropagates through the entire trajectory via a differentiable physics engine, providing exact gradients of the reward with respect to policy parameters. This leads to faster convergence and more stable policies, especially in high-frequency control tasks like agile interception.
The policy is a small feedforward neural network that takes as input: - The interceptor’s position (3), velocity (3), and orientation (4 as quaternion) - The intruder’s unit direction vector (3) relative to the interceptor
The output is four rotor thrust commands (scalars), which are fed into the quadrotor dynamics. During training, the simulator rolls out the policy for a fixed horizon, and the loss (negative reward) is backpropagated through time. The reward function encourages final proximity to the intruder and penalizes excessive control effort and high angular rates.
To handle the missing distance information, the policy implicitly learns to infer range from the direction vector’s rate of change and its own motion. This is a form of visual servoing without explicit depth estimation.
Why This Matters for Robotics
This work addresses a real bottleneck in autonomous pursuit: range estimation from monocular cameras is notoriously noisy and computationally expensive. By learning to intercept using only direction, the approach reduces sensor requirements to a single camera and an object detector, making it feasible for small, low-cost drones. Applications include drone swarm defense, search-and-rescue, and autonomous security patrols.
The use of differentiable simulation also points toward a broader trend in robotics: replacing sample-inefficient black-box RL with gradient-based training that leverages accurate physics models. This is particularly valuable for agile systems where every millisecond of latency matters. For fleet operators or integrators looking to deploy such capabilities, the open‑source release of the code lowers the barrier to experimentation.
If you're evaluating drones for security or interception tasks, check out drone options on Robot Overflow or explore used robotics platforms that could host these algorithms.
Limitations and Open Questions
The current approach assumes the intruder flies at a constant speed along a predefined trajectory—not an adaptive adversary. In a real adversarial setting, an evasive target would vary speed and direction in response to the pursuer, which the policy has not been tested against. Additionally, the work assumes the intruder stays within the interceptor’s field of view at all times and ignores sensor noise, occlusions, or detection failures.
Hardware validation remains an open milestone. All results are from simulation, and real-world deployment would introduce latency, imperfect actuation, and environmental disturbances not captured in the model. Future work aims to extend the policy to handle adversarial evaders and to integrate sensor uncertainty.
Frequently Asked Questions
What information does the policy need to intercept an intruder? Only the interceptor’s own state (position, velocity, orientation) and the unit direction vector pointing from the interceptor to the intruder—no distance or 3D position is required.
How is the policy trained? Using Analytic Policy Gradients (APG) through a differentiable quadrotor simulator, which backpropagates exact gradients of the reward through the full trajectory, enabling faster and more stable learning than traditional reinforcement learning.
Can this run on a real drone with a camera? Yes, the direction vector can be obtained from an onboard object detector (e.g., YOLO) using standard monocular cameras, making the approach lightweight and practical for small drones.
What are the main limitations? The intruder is assumed to move at constant speed on a non-evasive trajectory; the policy has not been tested on hardware; and sensor uncertainty or occlusions are not modeled.
Conclusion
By training a quadrotor interception policy in a differentiable simulator and relying only on directional information, this work demonstrates a practical path toward lightweight, camera-based pursuit drones. The combination of full-physics gradients and sparse sensing yields smoother and more reliable interception than conventional reinforcement learning, with clear next steps toward real-world deployment.