Robots that adapt online learn their own friction, drag, or payload effects while they run. They usually do this with a small neural network, a radial basis function (RBF) network, in which each neuron responds only to states near its own center. The network's output is added to the control command to cancel errors in the physical model. Because each neuron is local, a robot only learns about the parts of state space it has actually visited. When it enters a new region, the compensator there is blank, and the robot tracks poorly until errors grow large enough to retrain it. A fleet of similar robots collectively visits far more of state space than any single one. The question is how to pool what they have learned without streaming raw trajectories and without destabilizing each robot's fast control loop.
We propose that each robot periodically sends two things to a server: its RBF weights, and a compact record of how strongly it has excited each neuron. The server fuses the weights neuron by neuron, trusting each robot only where it has actually collected data. Robots do not overwrite their live weights with the result. Instead, their standard Lyapunov-derived update law gets a weak pull toward the fleet estimate, in place of the usual pull toward zero. The fleet estimate therefore acts as a moving anchor rather than a parameter jump. This makes the stability analysis simple, and it predicts that a robot's tracking error is bounded by how far its true dynamics sit from the fleet anchor.
Core Claim
Plain federated averaging of localized RBF weights is the wrong aggregation rule. It is what the closest prior work would suggest, since Federated Nonlinear System Identification runs FedAvg over linearly parameterized feature models. The rule fails because a robot that never visited a neuron's region reports a near-initial weight for it, and averaging lets those uninformed reports dilute the few informed ones. If only a fraction $\rho$ of robots excite a given neuron, averaging shrinks that weight toward its initial value by roughly a factor $\rho$.
Information-weighted fusion fixes this. Each robot $i$ keeps an excitation matrix $R_i=\int \lambda^{t-s}\Phi(x_s)\Phi(x_s)^\top ds$, where $\Phi$ is the vector of RBF activations and $\lambda<1$ is a forgetting factor. The server forms
$$\bar W = \Pi_{\mathcal B}\Big[\big(\textstyle\sum_i R_i + \mu I\big)^{-1}\textstyle\sum_i R_i W_i\Big],$$
where $\mu>0$ regularizes neurons nobody has visited and $\Pi_{\mathcal B}$ projects onto a known bounded set of plausible weights. Each robot then runs the modified adaptation law
$$\dot W_i = \Gamma\big(\Phi(x)\,e^\top P B - \sigma (W_i - \bar W)\big).$$
Here $e$ is tracking error, $P$ and $B$ come from the nominal error dynamics, $\Gamma$ is the adaptation gain, and $\sigma$ is the usual leakage ("sigma-modification") gain, now centered on $\bar W$ instead of zero.
The weights $W_i$ stay continuous when $\bar W$ changes, so the standard Lyapunov function $V=e^\top P e+\mathrm{tr}(\tilde W^\top\Gamma^{-1}\tilde W)$ never jumps. The usual sigma-modification argument then yields a uniformly ultimately bounded error. Its ultimate bound grows with the network's approximation error and with $\sigma\,\sup_t\|W_i^\star-\bar W(t)\|$, where $W_i^\star$ is robot $i$'s ideal weight vector. This bound has two parts. The heterogeneity floor $\epsilon$ that FedSysID and FNSysID prove is irreducible appears only through the anchor's prior-like pull. In directions a robot excites itself, its own adaptation dominates; in directions it never excites, it inherits the fleet's estimate. The gain to verify is therefore concentrated in transients and in unvisited regions, which is exactly where single-robot RBF control is known to fail.
Research Program
The work has three parts.
Analysis. We will prove the piecewise ultimate-boundedness result above for asynchronous, intermittent anchor updates, with the anchor confined to $\mathcal B$. We will then bound $\|W_i^\star-\bar W\|$ per neuron in terms of fleet-wide excitation and heterogeneity. The bound we expect to show is roughly $\epsilon$ plus a noise term shrinking as one over the square root of total fleet excitation at that neuron. This adapts the $1/\sqrt{M}$ argument of FNSysID from client count to per-neuron information. The FNSysID "small-ball" excitation condition is assumed fleet-wide rather than per client, which is the formal statement of "the fleet supplies excitation that individual missions lack."
Communication. For Gaussian RBFs on a grid, $R_i$ is nearly banded. We will test sending its diagonal only, a banded approximation, and 8-bit quantized versions. The quantization follows the compression in the AUV federated meta-transfer work. Transmissions will be event-triggered when a robot's excitation has grown enough to matter, borrowing the event-triggering logic of predefined-time event-triggered neural control.
Validation. We will run simulated fleets with controlled heterogeneity, and then, as a stretch, a small hardware fleet.
Two honest costs follow. First, $R_i$ reveals where each robot has operated. This weakens the privacy argument relative to sending weights alone, and we will report it rather than claim privacy. Second, if every robot excites the same neurons equally, the fusion rule reduces to FedAvg and the method offers nothing beyond FNSysID.
Baselines
Purely local adaptation. The naive reference is purely local RBF adaptive control with zero initial weights and standard sigma-modification toward zero. This is the controller family used in neuro-adaptive cooperative control and in the Baxter work in adaptive neural control with movement primitives, which reports errors within ±0.04 rad on its own setup. It is the strongest argument against federating at all, because it needs no communication. It will share identical RBF centers, gains, and nominal controller with our method. Beating it means lower tracking error during first entry into regions the robot has not visited, with no worse ultimate error.
FedAvg warm-start and FedAvg anchor. The strongest existing method is FNSysID-style FedAvg of the same RBF weights, used in two ways: to warm-start each robot's adaptive controller, and as an anchor with plain averaging. This is the obvious combination of the two fields and the closest competitor. On Lacuna, FNSysID reports only the $1/\sqrt{M}$ scaling of parameter error with no absolute value. FedSysID reports about a 10× parameter-error reduction at $M=100$ versus $M=1$ on a linear system. No reported closed-loop tracking result was found on Lacuna for federated RBF control. It gets the same communication schedule and the same number of rounds. Beating it means a statistically clear tracking-error reduction in heterogeneous-coverage regimes, and parity when coverage is uniform.
Centralized pooling. Raw trajectories from all robots are pooled and fit offline by least squares. This is an upper bound on accuracy and shows how much our fusion loses. We will report its uplink cost.
Classical robust control. A sliding mode controller with a boundary layer, tuned to the known disturbance bound, gives a no-learning reference for tracking error and chattering.
Ablations. Count-weighted fusion uses only the diagonal of $R_i$. A hard-reset variant overwrites $W_i\leftarrow\bar W$ at each round instead of anchoring. The first ablation isolates whether the off-diagonal excitation structure matters. The second isolates whether the anchoring mechanism, rather than the fusion rule, is what prevents transients.
Why not just…?
Why not just warm-start each robot from FedAvg weights and let local adaptation fix the rest? This is the main threat, because local adaptation does remove the heterogeneity floor in regions the robot excites. Its weakness is specific. Averaged weights for a sparsely visited neuron are diluted by robots that never visited it, so the warm start is weakest exactly where it is most needed. The coverage-split experiment below measures this directly. If plain averaging matches information-weighted fusion there, the proposal reduces to FNSysID plus adaptive control, and we will say so.
Why not just stream trajectories and fit centrally? Centralized fitting is the accuracy ceiling, and we run it. The cost is bandwidth: the AUV work reports large savings from sending model pieces instead of raw streams, including 4.5× information gain per Joule over centralized streaming (AUV FMTL). Sending $W_i$ and a banded $R_i$ costs on the order of the number of neurons per round, independent of trajectory length. We will report the accuracy gap against bytes sent.
Why not just use sliding mode control or a disturbance observer? These guarantee robustness but pay for it with chattering or high gains sized to worst-case bounds. Learned compensation can avoid that cost, as in the 10⁻⁴ versus 10⁻²–10⁻³ error scales reported for friction compensation in KUKA friction estimation. We measure high-frequency control energy alongside tracking error, so a fair comparison is visible rather than assumed.
Why not just weight robots by how many samples they collected? If neurons barely overlap, per-neuron count weighting is nearly the same as our rule. The diagonal-only ablation tests this. If it matches the full rule, we will adopt the simpler version.
Proposed Experiments
Platforms. The main platform is a simulated quadrotor fleet in safe-control-gym and PyBullet. Each of $M\in\{1,5,20,100\}$ quadrotors gets a mass and drag perturbation drawn within a controlled heterogeneity radius $\epsilon$, plus wind gusts. The RBF network compensates the residual force. The second platform is a simulated 7-DOF manipulator with per-joint Stribeck friction parameters varied across the fleet, modeled on the KUKA iiwa setting.
Metrics. We record RMS tracking error over the first pass through each region, ultimate tracking error, per-neuron weight error against ground truth, high-frequency control energy as a measure of chattering, and uplink bytes.
Decisive check: coverage split. Each robot flies missions confined to a different slice of the workspace, with overlap fraction $\rho$ varied from 1 down to $1/M$. Every robot is then tested on trajectories spanning the whole workspace. The prediction is that FedAvg-anchored tracking error in unvisited regions degrades as $\rho$ falls, while information-weighted fusion stays near the centralized ceiling. At $\rho=1$ the two should coincide.
Second check: heterogeneity sweep. We sweep $\epsilon$ at fixed coverage. Per-neuron weight error should plateau at a level set by $\epsilon$, as FedSysID and FNSysID predict. Tracking error in visited regions should not plateau, because local adaptation absorbs the residual.
Third check: reset versus anchor. We compare hard reset against anchoring under rounds that are asynchronous and arrive every few seconds. The hard-reset variant should show tracking spikes at aggregation instants, and the anchored version should not.
Hardware stretch. If simulation results hold, the stretch goal is a small Crazyflie 2.1 fleet. The DATT paper (DATT) reports 0.05 m star-trajectory error for its learned controller versus 0.11 m for MPC on single vehicles. That setup differs, so it is a reference point for scale rather than a target.
References
- Federated Nonlinear System Identification
- FedSysID: A Federated Approach to Sample-Efficient System Identification
- System-Level Optimization of AUV Swarm Control and Perception: An Energy-Aware Federated Meta-Transfer Learning Framework with Digital Twin Validation
- Predefined-Time Adaptive Neural Tracking Control for a Single Link Manipulator with an Event-Triggered Mechanism
- Neuro-adaptive cooperative control for high-order nonlinear multi-agent systems with uncertainties
- Robot Learning System Based on Adaptive Neural Control and Dynamic Movement Primitives
- Adaptive Control Based Friction Estimation for Tracking Control of Robot Manipulators
- Safe-Control-Gym: A Unified Benchmark Suite for Safe Learning-Based Control and Reinforcement Learning in Robotics
- DATT: Deep Adaptive Trajectory Tracking for Quadrotor Control