💻 computer science

RVC-NMPC: Nonlinear Model Predictive Control with Reciprocal Velocity Constraints for Mutual Collision Avoidance in Agile UAV Flight

This paper introduces RVC-NMPC, a high-frequency (100 Hz) nonlinear model predictive control framework that integrates time-dependent reciprocal velocity constraints using only observable data to enable safe, agile mutual collision avoidance for UAVs in dense, high-speed scenarios without requiring excessive communication.

Original authors: Vit Kratky, Robert Penicka, Parakh M. Gupta, Ondrej Prochazka, Martin Saska

Published 2026-07-30

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Vit Kratky, Robert Penicka, Parakh M. Gupta, Ondrej Prochazka, Martin Saska

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy sky, not filled with clouds, but with a swarm of tiny, buzzing drones. For years, these flying robots were mostly solo artists, working alone in empty rooms. But now, we want them to work together in the real world—delivering packages, patrolling cities, and weaving through traffic. The big problem? If they all fly fast and close together, they might crash. To stop this, scientists use a mix of "brain" and "reflexes." The "brain" is a planning system that figures out the best path to a destination. The "reflexes" are safety rules that kick in when another robot gets too close, telling the drone to swerve or slow down instantly. The challenge is making these drones fast enough to be useful but smart enough to never bump into each other, all while using only the information they can see or hear right now, without needing a super-fast internet connection to talk to every other drone.

This paper introduces a new way to fly these drones called RVC-NMPC. Think of it as teaching a drone to play a high-speed game of "chicken" with its friends, but with a twist: instead of just guessing what the other player will do, it calculates a "safe zone" based on where they are right now and how fast they are moving. The authors call this a "Reciprocal Velocity Constraint." In plain English, it's like a shared agreement between two drones: "If I keep going this way, and you keep going that way, we'll crash. So, let's both nudge our paths just a little bit to stay safe." Unlike older methods that try to predict the entire future flight path of every other drone (which requires a lot of talking and computing power), this new method only looks at the current speed and position. It then feeds this safety rule directly into the drone's main flight computer, which uses complex math to figure out exactly how to move its motors to stay safe while still flying super fast.

The researchers tested this idea in two ways: first, in a detailed computer simulation, and second, with real drones flying in a lab. In the simulation, they put 10 drones in a tight circle, all trying to fly to the opposite side at the same time. The new method was incredibly fast, completing the task 31% quicker than the best existing methods, while keeping all the drones from crashing. They even tested it with real drones flying at speeds up to 18 meters per second (about 40 mph) and accelerating as hard as 30 meters per second squared. The system ran so efficiently that it could make decisions 100 times every second. This means the drones could react instantly to changes, like a friend suddenly changing direction. The paper shows that by using these time-dependent safety rules, the drones can fly much more aggressively and efficiently than before, without needing to know the future plans of their neighbors. While the method doesn't have a perfect mathematical guarantee that it will never fail in every possible weird situation, the tests show it is extremely reliable in real-world-like scenarios, handling delays in communication and even a bit of "noise" in the data without crashing.

Technical Summary: RVC-NMPC for Mutual Collision Avoidance in Agile UAV Flight

Problem Statement
The rapid deployment of Uncrewed Aerial Vehicles (UAVs) for tasks such as package delivery and area monitoring necessitates operation in dense, shared airspace. While centralized planning is often impractical for high-speed scenarios due to communication delays and scalability issues, existing decentralized methods face significant limitations. Current state-of-the-art approaches typically require the exchange of full future trajectories, imposing high bandwidth demands, or rely on simplified kinematic models that neglect the nonlinear dynamics of UAVs. Furthermore, many existing methods are restricted to low velocities (often below 10 m/s), failing to leverage the agility of commercial drones capable of higher speeds and accelerations. There is a critical need for a decentralized collision avoidance strategy that operates at high frequencies, accounts for full nonlinear dynamics, and relies only on observable current states (position and velocity) rather than predicted future paths.

Methodology
The authors propose RVC-NMPC, a framework integrating Nonlinear Model Predictive Control (NMPC) with time-dependent Reciprocal Velocity Constraints (RVCs). The system architecture consists of a pipeline running at 100 Hz on an ARM processor, comprising:

State Estimation: Processes sensor data to estimate the UAV's current position and velocity.
Reference Trajectory Generation: Uses a Point-Mass Model (PMM) to generate a minimum-time reference trajectory to the goal, respecting kinematic constraints but ignoring collisions.
Reciprocal Velocity Constraint Generator: Computes linear constraints based on the current positions and velocities of neighboring robots. Unlike methods that compute constraints for every point in the prediction horizon, this module generates a single set of constraints derived from the current state.
NMPC Controller: Solves a nonlinear optimization problem to generate control inputs (motor thrusts). The objective minimizes the deviation from the reference trajectory and control effort. Crucially, the collision avoidance is enforced via soft linear constraints ( $A_{m,k}v_k + s_{m,k} \ge b_{m,k}$ ) integrated directly into the NMPC formulation.

Key Technical Innovations

Time-Dependent Constraints: The paper introduces a novel mechanism where reciprocal velocity constraints are valid only for a specific time window ( $t_{v,m}$ ). This duration is calculated based on the relative position and velocity vectors; the constraint is active only until the relative velocity vector would cause the robots to pass each other (i.e., the angle between relative position and relative velocity exceeds $\pi/2$ ). This prevents the constraints from being overly restrictive for the entire prediction horizon, allowing the UAV to maneuver more freely once the immediate collision risk has passed.
Single-Set Constraint Application: Instead of computing unique constraints for every transition point in the NMPC horizon (which increases computational complexity), the method applies a single set of constraints over the entire horizon. Due to the convexity of the feasible velocity set, any convex combination of feasible velocities remains feasible. This approach drastically reduces computational load while maintaining collision avoidance guarantees.
Direct Integration of Dynamics: The controller utilizes a full nonlinear quadrotor dynamic model (including body rates and quaternion orientation) rather than simplified kinematic models, enabling agile flight at high accelerations.
Low-Bandwidth Operation: The method requires only the current position and velocity of other agents, obtainable via onboard sensing or low-bandwidth communication (e.g., Remote Drone ID), eliminating the need for high-bandwidth trajectory sharing.

Results
The approach was evaluated through extensive simulations and real-world experiments:

Simulation Performance: In a challenging scenario involving 10 UAVs navigating to antipodal positions on a 10m radius circle with velocity limits of 20 m/s and acceleration limits of 40 m/s, RVC-NMPC achieved a 31% reduction in flight time compared to the best-performing state-of-the-art method (HDSM), while maintaining a 100% success rate (collision-free). It outperformed methods like MADER, EGO-SWARM-2, and RBL.
Robustness: The system demonstrated resilience to communication delays up to 50 ms and message frequencies down to 10 Hz. It also maintained performance with state estimation noise up to $\sigma_p = 1$ m and $\sigma_v = 2$ m/s.
Reliability: A 3-hour continuous simulation with 10 UAVs navigating to 50,000 random goals with accelerations up to 40 m/s resulted in zero collisions.
Real-World Validation: Experiments with three UAVs on a Khadas VIM3 processor (2 GHz ARM) confirmed the approach's viability. The system successfully navigated with accelerations up to 30 m/s² and velocities up to 18 m/s, utilizing a 10 Hz communication link with an average delay of 14 ms.

Significance and Claims
The paper claims that RVC-NMPC represents a significant advancement for high-speed, agile UAV flight in shared environments. Its primary significance lies in its ability to combine full nonlinear dynamic modeling with real-time performance (100 Hz) without requiring the exchange of future trajectories. By relying solely on observable current states and employing time-dependent constraints, the method achieves superior efficiency (reduced flight time) and robustness compared to existing ORCA-MPC or trajectory-sharing approaches. The authors emphasize that while the method lacks theoretical guarantees (due to the absence of future trajectory knowledge and the use of soft constraints), its practical reliability has been demonstrated through extensive testing in scenarios exceeding the speed and acceleration limits of most existing literature. The approach is presented as a practical solution for real-world deployment where communication bandwidth is limited and dynamic agility is required.

More like this