MetaTune: Adjoint-based Meta-tuning via Robotic Differentiable Dynamics

Uncertainties and disturbances in robotic systems, such as aerodynamic forces, are fundamentally outcomes of physical interactions with the environment, manifesting as learnable spatiotemporal sequences rather than random noise. However, achieving high-precision control for robotic systems operating in unstructured environments is often hindered by complex unmodeled dynamics and external disturbances. While learning-based methods offer powerful approximation capabilities, they typically suffer from heavy reliance on offline training and lack theoretical guarantees. Conversely, traditional robust control strategies are predominantly reactive, limited to instantaneous estimation without the foresight to anticipate future disturbance trends. To bridge this gap, this paper proposes a differentiable data-enabled Koopman control framework termed DEKC. Unlike black-box approaches, DEKC adopts a hybrid modeling strategy that retains the nominal physics model while employing a deep neural network to parameterize the lifting function of Koopman operator for unknown residual dynamics. Crucially, the framework formulates disturbances as a dynamical system, learning their temporal evolution in a global linear space. This enables the prediction of future disturbance trajectories, which are explicitly integrated into controller for preemptive compensation. Furthermore, an online backward gradient update mechanism is introduced to ensure real-time adaptation to time-varying uncertainties. Numerical simulations on a tethered space robot demonstrate the efficacy of the proposed DEKC in mitigating highly coupled uncertainties. Complementing these results, real-world experiments on a quadrotor substantiate its superiority in tracking agile trajectories under uncertainties induced by aerodynamics and suspended payload.

In this paper, we address the problem of reference tracking for uncertain nonlinear systems. Since collecting data from the target system (i.e., the system of interest) is often challenging, our objective is to design optimal controllers using limited target system data. Meta-learning provides a promising paradigm by leveraging offline data from source systems (systems sharing structural similarities with the target system) to accelerate training and enhance control performance. Motivated by this idea, we propose a meta-learning-based control framework that tailors the implicit model-agnostic meta-learning (iMAML) algorithm to the control setting. The framework operates in two phases: an (offline) meta-training phase, where an aggregated representation is learned from source data to capture the shared system dynamics among similar systems, and an (online) meta-adaptation phase, where this representation is fine-tuned on the target system using only a few data samples and limited adaptation steps. We formulate this framework as a bi-level optimization problem and provide an efficient solution with reduced storage complexity and few approximations. The proposed framework is general, allowing various learning algorithms to be integrated. To demonstrate this flexibility, we propose two specific learning algorithms that can be incorporated into our framework based on a neural state-space model and a deep Q-network, respectively. The primary distinction between these approaches is whether explicit system identification is required. Numerical simulations and hardware experiments demonstrate that the proposed methods enhance control performance and consistently outperform baseline approaches.

Adaptive control of Euler-Lagrange systems is challenging when friction is governed by a finite-horizon internal state that is not directly observable from joint measurements. In this setting, the measured closed-loop state is no longer Markovian, and standard certainty-equivalence adaptive laws may lose their convergence guarantees. The paper proposes a meta-control architecture in which the gains of a computed-torque controller are generated by a self-attention block processing a short window of recent motion history. The number of attention heads is selected before policy training through a surrogate analysis of the autocovariance of the memory-state gradient along the temporal window. This surrogate is based on a temporal adaptation of an incremental rank-tracking framework previously developed by the authors. The selected head count is then fixed and used as an architectural hyperparameter in a reinforcement-learning stage, where the policy is trained under a shielded admissibility constraint. The approach is tested on a 2-DOF manipulator with nonlinear friction and variable payload. In the short and matched memory regimes, the single-layer attention-only meta-controller outperforms a deeper Transformer baseline, with tracking-error reductions of 12 and 19 percentage points, respectively. The reported effect sizes are large, with d approximately -1.1 and -2.1, and Mann-Whitney p < 0.05 in both cases. In the long memory regime, however, the advantage disappears. Four out of ten training runs show either divergence or payload-invariant policy collapse, revealing a weakness in the static Phase-1 head-count prescription. This motivates moving rank-tracking inside the reinforcement-learning loop, allowing attention heads to be pruned or grown at runtime instead of fixed before training.