KOALA: Koopman Operator Learning for WiFi-Based Anticipatory Human Motion Prediction

Anticipating human motion from ambient radio-frequency signals is crucial for privacy-preserving smart environments, healthcare monitoring, and human-robot collaboration. While WiFi Channel State Information (CSI) can estimate instantaneous 3D poses without cameras, forecasting future motion remains challenging. Existing approaches treat inference as an instantaneous regression problem, and naively applying vision-based prediction models to CSI fails because autoregressive rollouts amplify estimation noise and spatial jitter at every step. This paper asks whether human motion can be predicted directly from WiFi CSI across multiple time horizons without suffering from error accumulation.

The authors propose KOALA (Koopman Operator with Attentive Lifting Architecture), a framework that processes raw WiFi CSI through four sequential stages. First, a Mamba-based CSI encoder extracts per-frame features and a global context vector. Second, a joint-conditioned human pose estimation module recovers 3D joint coordinates, which are then mapped by a skeleton-aware encoder into kinematically structured features while blocking gradients to isolate geometric accuracy. Third, a dual-stream fusion combines CSI and pose features before a lifting network projects them into a higher-dimensional Koopman latent space. Within this latent space, a residual CSI-conditioned operator evolves the state across multiple future horizons using single matrix-vector products, avoiding autoregression. Finally, an anchor-based decoding head predicts pose deltas relative to an anchor pose.

KOALA introduces several specialized components to overcome limitations in prior Koopman and WiFi-based sensing literature. The residual CSI-conditioned operator resolves the identity attractor problem where naive linear operators collapse to trivial identity mappings. Additionally, an anchor-delta prediction formulation prevents the model from taking the degenerate shortcut of copying the current pose across all horizons. To jointly regularize the lifting and operator without separate contrastive or spectral objectives, the authors introduce a Koopman Anchored Latent loss operating directly in the temporal-encoder feature space.