Primitive-Swarm: An Ultra-lightweight and Scalable Planner for Large-scale Aerial Swarms

TL;DR

Primitive-Swarm delivers an ultra-lightweight, scalable planner for large-scale aerial swarms by decoupling online planning into linear-complexity primitive selection. It offline-generates a time-optimal motion primitive library via TOPP-RA and precomputes spatial/spatio-temporal occupancy relationships to enable batched collision checks on raw point clouds, avoiding map inflation. The system operates in a decentralized, asynchronous fashion with a velocity-aligned frame and receding-horizon replanning, achieving real-time planning in dense environments and scaling to 1000-agent swarms, including real-world SWaP-constrained flights. The combination of offline TOPP-RA primitives, offline occupancy indexing, and online linear-cost selection yields strong performance gains in flight time, path length, and computation time, while maintaining robust safety guarantees in unknown environments.

Abstract

Achieving large-scale aerial swarms is challenging due to the inherent contradictions in balancing computational efficiency and scalability. This paper introduces Primitive-Swarm, an ultra-lightweight and scalable planner designed specifically for large-scale autonomous aerial swarms. The proposed approach adopts a decentralized and asynchronous replanning strategy. Within it is a novel motion primitive library consisting of time-optimal and dynamically feasible trajectories. They are generated utlizing a novel time-optimial path parameterization algorithm based on reachability analysis (TOPP-RA). Then, a rapid collision checking mechanism is developed by associating the motion primitives with the discrete surrounding space according to conflicts. By considering both spatial and temporal conflicts, the mechanism handles robot-obstacle and robot-robot collisions simultaneously. Then, during a replanning process, each robot selects the safe and minimum cost trajectory from the library based on user-defined requirements. Both the time-optimal motion primitive library and the occupancy information are computed offline, turning a time-consuming optimization problem into a linear-complexity selection problem. This enables the planner to comprehensively explore the non-convex, discontinuous 3-D safe space filled with numerous obstacles and robots, effectively identifying the best hidden path. Benchmark comparisons demonstrate that our method achieves the shortest flight time and traveled distance with a computation time of less than 1 ms in dense environments. Super large-scale swarm simulations, involving up to 1000 robots, running in real-time, verify the scalability of our method. Real-world experiments validate the feasibility and robustness of our approach. The code will be released to foster community collaboration.

Figures (23)

• Figure 1: Large-scale (1000 drones) air traffic simulation in an unknown environment. The green ellipsoids represent the drones, while the gray pillars represent skyscrapers serving as obstacles. The colored curves illustrate the executed trajectories of the drones. Drones start from five layers L1 to L5, as shown in Figure (a), and then fly to layers L4, L5, L3, L1, L2, respectively. It is notable that all drones operate independently in separate threads, following a decentralized and asynchronous architecture.
• Figure 2: Qualitative comparison with RBPpark2020efficient, MADERtordesillas2021mader, EGO-Swarmzhou2021ego, MINCO-Swarmzhou2022swarm, and AMSwarmXadajania2023amswarmx. Axis ticks from inside to outside represent: Computation efficiency - heavyweight, moderate, lightweight, ultra lightweight; Scalability - low, medium, high; Trajectory quality - low, medium, high, very high; System Autonomy - none, partially autonomous, fully autonomous. Please refer to Section \ref{['sec:RW_trajectory_planning']} for the detailed discussion of the qualitative comparison.
• Figure 3: The overview of our decentralized and asynchronous autonomous aerial swarm system, which includes state estimation (yellow), planning (blue), control (green), and communication (red) modules. $\mathcal{W}$ refers to the world coordinate system and $\mathcal{V}$ is called the velocity-aligned frame defined by current drone position and velocity in Sec. \ref{['sec:vcs']}.
• Figure 4: An example of path library generation (Note: Path refers to a geometric curve without temporal information). (a) Seven orange arcs representing paths; (b) Front view of the path library obtained by rotating the arcs in the sub-figure (a) around the x-axis with an angle interpolation of $D_{angle}=30^\circ$; (c) Side view of the path library, where both blue and orange arcs represent paths. The origin of all paths coincides with the origin of the velocity-aligned coordinate system $\mathcal{V}$. All paths start tangent to the x-axis. The orange and red dots represent the end points of each path.
• Figure 5: Time parameterization process for path $\mathbf{q}(s)$ with specified start and end velocities, $\dot{s}_0$ and $\dot{s}_N$. Step 1 involves discretizing $s$. In Step 2, dynamical constraints are applied to each discrete position $s_i$. Step 3 encompasses the sequential computation of the controllable set $\mathcal{K}_i(\mathbb{I}_N)$ (illustrated as blue intervals) in a backward manner, starting from the discrete position $s_N$. In Step 4, we iteratively select the largest admissible control $u_i^*$ from the discrete position $s_0$ in a forward manner. The optimal state $x_{i+1}^* := x_i^* + 2\Delta_iu_i^*$ (represented as red points) ensures that it remains within the corresponding controllable set $\mathcal{K}_{i+1}(\mathbb{I}_N)$.