Ghost Attractor Networks: Efficient Dynamical Decoder for Robotic Control

Researchers have introduced Ghost Attractor Networks, a novel machine learning approach designed to improve sequential output generation for robotic control systems. The method addresses fundamental efficiency challenges inherent in large-scale Transformer and diffusion decoders, which incur memory costs that scale with sequence length and require iterative per-step computation.

Ghost Attractor Networks employ a theoretically grounded dynamical decoder architecture whose latent representations evolve under a learned potential with drift, producing a basin-attractor structure by design. This approach enables three critical capabilities: multi-modality handling, decoder-level single-pass switching, and constant memory consumption. The network operates through mode transitions modeled as saddle-node bifurcations with ghost-attractor escape dynamics. A hierarchical phase-space decomposition separates first-order basin convergence from second-order proprioceptive refinement, allowing both high-level mode selection and fine-grained action generation.

Empirical validation demonstrates substantial efficiency gains. A 2.3-million-parameter Ghost network achieves offline accuracy matching a 1.07-billion-parameter Diffusion Transformer while requiring 462 times fewer parameters and delivering 32 times lower latency. The model outperforms five alternative 2-million-parameter baselines—including MLP, Neural ODE, CVAE, Transformer, and 1-step Diffusion decoders—on offline mean squared error by margins ranging from 5.9 to 29 percent. Training with behavioral-cloning and contrastive objectives produces the theoretically predicted gradient-flow contraction, with gradient norm decaying by 67 percent across five integration steps on 1,430 held-out test samples.

In closed-loop robotic control benchmarks, Ghost networks deliver substantial performance improvements. Phase conditioning on the basin-structured latent representation yields a 13.5 percentage-point success-rate advantage over baseline feed-forward MLP approaches on the LIBERO-10 benchmark. Persistent-latent ensembling further elevates performance to a 95.7 percent final success rate, demonstrating the robustness of the basin-structured latent geometry for maintaining stable multi-step robot action sequences.