MeanFlow Meets Control: Scaling Sampled-Data Control for Swarms
| Authors | Anqi Dong et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2603.20189 |
| Download | |
| Categories | cs.LG, cs.MA, cs.RO |
Abstract
Steering large-scale swarms in only a few control updates is challenging because real systems operate in sampled-data form: control inputs are updated intermittently and applied over finite intervals. In this regime, the natural object is not an instantaneous velocity field, but a finite-window control quantity that captures the system response over each sampling interval. Inspired by MeanFlow, we introduce a control-space learning framework for swarm steering under linear time-invariant dynamics. The learned object is the coefficient that parameterizes the finite-horizon minimum-energy control over each interval. We show that this coefficient admits both an integral representation and a local differential identity along bridge trajectories, which leads to a simple stop-gradient training objective. At implementation time, the learned coefficient is used directly in sampled-data updates, so the prescribed dynamics and actuation map are respected by construction. The resulting framework provides a scalable approach to few-step swarm steering that is consistent with the sampled-data structure of real control systems.
Engineering Breakdown
Plain English
This paper addresses the challenge of controlling large swarms of agents when control updates can only happen intermittently—a realistic constraint in real systems where you can't update commands every microsecond. Rather than working with instantaneous velocity fields (the traditional approach), the authors propose learning a control coefficient that represents the optimal finite-horizon control action over each sampling interval. They frame this as a control-space learning problem under linear time-invariant dynamics, showing that the learned coefficient can be computed efficiently using both integral representations and differential identities. The key innovation is that this approach scales to large swarms while requiring only a few control updates, making it practical for real distributed systems.
Core Technical Contribution
The core novelty is reframing swarm control from continuous instantaneous commands to sampled-data control by learning a finite-window control coefficient parameterization. This coefficient captures the minimum-energy control action over each sampling interval, which is fundamentally different from prior MeanFlow approaches that work with velocity fields at single time points. The authors prove this coefficient admits both an integral representation (exact computation) and a local differential identity (enabling efficient gradient-based learning with stop-gradient optimization). This dual representation is the key technical insight—it allows them to train control policies that respect the actual discrete-time, finite-interval nature of real hardware while maintaining computational efficiency.
How It Works
The system takes as input the current state of a swarm and uses learned control coefficients to determine what finite-horizon minimum-energy control action to apply over the next sampling interval. Rather than learning a continuous policy π(x) that outputs instantaneous velocities, they learn a coefficient that parameterizes the solution to a finite-horizon optimal control problem—essentially learning to query the right solution from a pre-computed family of controls. The integral representation allows exact computation of the control coefficient by integrating over the sampling window, while the differential identity enables backpropagation-friendly training without explicitly solving the optimal control problem each forward pass. During training, they use stop-gradient operations on certain computations to stabilize learning—essentially treating some intermediate values as constants rather than differentiable paths. The learned coefficients are applied over discrete sampling intervals, and multiple intervals of control updates steer the swarm toward desired configurations.
Production Impact
For engineers building swarm robotics systems, drone fleets, or distributed multi-agent systems, this directly addresses the gap between theory (continuous control) and reality (discrete hardware updates). You would replace traditional continuous control laws or discrete MPC solvers with pre-learned control coefficients that execute in real-time with minimal computation—crucial when you have hundreds or thousands of agents with limited onboard compute. The sampled-data framing means your control stack maps cleanly to actual hardware timing: compute the coefficient, apply it for the fixed sampling period (e.g., 10ms), repeat. This eliminates the common problem of control algorithms assuming infinitesimal updates that can't actually be implemented. The trade-off is upfront: you need offline training to learn the coefficient parameterization, but once trained, inference is extremely cheap compared to online optimization methods like MPC. Integration is straightforward in any robotics stack—replace your control law function with a learned coefficient lookup.
Limitations and When Not to Use This
This approach assumes linear time-invariant (LTI) dynamics, which excludes many real swarm scenarios with nonlinear interactions, friction, collision avoidance, or communication delays—you'd need extensions to handle these practical complexities. The paper doesn't discuss how performance degrades when the actual swarm dynamics deviate from the LTI assumptions used during training, which is critical for safety-critical applications. It also doesn't address heterogeneous swarms where agents have different dynamics, or when the swarm size changes dynamically—the scalability claim needs validation on orders of magnitude larger than typical experimental swarms. The finite-window formulation requires knowing or estimating the sampling interval and system parameters precisely; robustness to parameter uncertainty and sampling jitter isn't thoroughly explored. Finally, the paper appears incomplete in the abstract (cuts off at 'o'), so the full experimental validation, sample complexity requirements, and convergence guarantees cannot be fully evaluated.
Research Context
This work builds on the MeanFlow framework, which uses integral representations of control flows, but extends it to the discrete-time sampled-data regime that actual hardware requires. It sits at the intersection of control theory (finite-horizon optimal control, LTI systems) and machine learning (learning to parameterize solutions rather than solving them online). The contribution is significant because most deep learning approaches to swarm control ignore the sampling constraint or assume continuous updates, while classical control theory handles sampling but doesn't leverage learning—this paper bridges that gap. The research direction opens up applying learned control parameterizations to other sampled systems in robotics, autonomous vehicles, and distributed optimization, potentially making control-theoretic guarantees compatible with learned policies at scale.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::