Neural-Swarm2: Planning and Control of Heterogeneous Multirotor Swarms Using Learned Interactions

Abstract

We present Neural-Swarm2, a learning-based method for motion planning and control that allows heterogeneous multirotors in a swarm to safely fly in close proximity. Such operation for drones is challenging due to complex aerodynamic interaction forces, such as downwash generated by nearby drones and ground effect. Conventional planning and control methods neglect capturing these interaction forces, resulting in sparse swarm configuration during flight. Our approach combines a physics-based nominal dynamics model with learned Deep Neural Networks (DNNs) with strong Lipschitz properties. We evolve two techniques to accurately predict the aerodynamic interactions between heterogeneous multirotors: i) spectral normalization for stability and generalization guarantees of unseen data and ii) heterogeneous deep sets for supporting any number of heterogeneous neighbors in a permutation-invariant manner without reducing expressiveness. The learned residual dynamics benefit both the proposed interaction-aware multi-robot motion planning and the nonlinear tracking control designs because the learned interaction forces reduce the modelling errors. Experimental results demonstrate that Neural-Swarm2 is able to generalize to larger swarms beyond training cases and significantly outperforms a baseline nonlinear tracking controller with up to three times reduction in worst-case tracking errors. Video is available at https://youtu.be/Y02juH6BDxo.

Neural-Fly enables rapid learning for agile flight in strong winds

“…Nonlinear Control Law We start by defining some notation. The composite velocity tracking error term s and the reference velocity qr are defined such that s = q − qr = q + Λq (10) where q = q − q d is the position tracking error and Λ is a positive definite gain matrix. Note when s exponentially converges to an error ball around 0, q will exponentially converge to a proportionate error ball around the desired trajectory q d (t) (see Section S5).…”

“…(ii) The normalization (line 6) is to make sure ‖a*‖ ≤ , which improves the robustness of our adaptive control design. We also use spectral normalization in training , to control the Lipschitz property of the neural network and improve generalizability (8,10,12). (iii) We train h and  in an alternating manner.…”

“…Recent state of the art in quadrotor flight control has used adaptive control methods that directly estimate the unknown aerodynamic force without assuming the structure of the underlying physics, but relying on high-frequency and low-latency control (4)(5)(6)(7). In parallel, there has been increased interest in data-driven modeling of aerodynamics [e.g., (8)(9)(10)(11)]; however, existing approaches cannot effectively adapt in changing or unknown environments, such as time-varying wind conditions.…”

Decentralized Control of Quadrotor Swarms with End-to-end Deep Reinforcement Learning

“…[12] and [13] both utilize Imitation Learning (IL) from a global centralized planner, while [13] further utilizes RL to balance the tradeoff between actions optimal to the team and individuals. [14,15] learn controllers robust to aerodynamic interactions such as downwash. Graph Neural Networks (GNNs) are promising architectures for swarm control and they have been proposed as a parameterization for imitation learning [16,17] and RL [18,19] algorithms.…”

Online Optimization with Feedback Delay and Nonlinear Switching Cost

“…The impact of multi-round delay has been recognized as a challenging open question for the design of online algorithms and broadly in applications. For example, [3] highlights that a three-step delay (30 milliseconds) can already cause catastrophic crashes in drone tracking control using a standard controller without algorithmic adjustments for delay.…”