Continual learning based state-aware estimation for reliable industrial legged robots

Reliable state estimation is fundamental to learning-based control of legged robots. Existing learning-based locomotion estimators typically predict only limited privileged states or learn implicit latent representations, while insufficiently addressing the non-stationary data distribution induced by evolving reinforcement-learning policies. In this setting, explicit multi-state estimation involves three coupled challenges: scalable estimation of heterogeneous robot states under onboard constraints, task-aware utilization of estimated states, and preservation of estimator plasticity during non-stationary training. To address these challenges, we propose SENSE, a state-aware estimation framework for reinforcement-learning-based legged locomotion. SENSE employs a sparse Mixture-of-Experts State Estimation Module (SEM) to infer diverse explicit robot states from proprioceptive histories, and introduces a StateSelector to forward task-relevant estimates to the control policy. We further propose CBP(COMP), a competitive continual backpropagation mechanism that replaces random neuron reinitialization with competitive parameter inheritance to alleviate unbalanced neuron turnover and preserve plasticity. Mechanism-level analysis explains its asymptotic competition symmetry under simplified assumptions. Controlled estimator comparisons, ablations, reset-distribution statistics, dynamic-payload adaptation, and real-world hexapod experiments show improved estimation accuracy, control performance, continual adaptation, and deployment robustness. These results suggest a practical neural state-estimation paradigm for interpretable, adaptive, and reliable learning-based legged robots.