Tiny Recursive Control: Iterative Reasoning for Efficient Optimal Control

2025-12-18

AI Summary

To address the high power consumption and latency incurred by large-parameter neural controllers in embedded aerospace systems, this paper proposes a novel optimal control paradigm that recursively deepens rather than widens the network. The core innovation is the first integration of recursive inference into continuous optimal control synthesis: a compact, fixed-size neural network (1.5M parameters) is reused across multiple inference rounds, augmented by dual-layer latent-space modeling, trajectory simulation, and tracking-error-driven iterative refinement—enabling emergent representational capacity. Computation scales with iteration count, while memory footprint remains constant (<10 MB). Evaluated on oscillator stabilization and fuel-constrained powered descent tasks, the method achieves near-optimal control performance with millisecond-scale GPU inference latency. Compared to large language model–based baselines, it reduces both parameter count and memory usage by two orders of magnitude.

Technology Category

• Search and Optimization: Learning to Search
• Natural Language Processing: Learning & Optimization for NLP
• Intelligent Robots: Learning & Optimization for ROB

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Neural network controllers increasingly demand millions of parameters, and language model approaches push into the billions. For embedded aerospace systems with strict power and latency constraints, this scaling is prohibitive. We present Tiny Recursive Control (TRC), a neural architecture based on a counterintuitive principle: capacity can emerge from iteration depth rather than parameter count. TRC applies compact networks (approximately 1.5M parameters) repeatedly through a two-level hierarchical latent structure, refining control sequences by simulating trajectories and correcting based on tracking error. Because the same weights process every refinement step, adding iterations increases computation without increasing memory. We evaluate TRC on nonlinear control problems including oscillator stabilization and powered descent with fuel constraints. Across these domains, TRC achieves near-optimal control costs while requiring only millisecond-scale inference on GPU and under 10~MB memory, two orders of magnitude smaller than language model baselines. These results demonstrate that recursive reasoning, previously confined to discrete tasks, transfers effectively to continuous control synthesis.

Problem

• Addresses prohibitive scaling of large neural controllers for embedded aerospace systems.
• Develops a compact neural architecture using iteration depth instead of parameters.
• Enables near-optimal control with minimal memory and millisecond inference times.

Innovation

• Recursive neural architecture with iterative refinement
• Hierarchical latent structure for trajectory simulation
• Compact network reused across refinement steps