Tiny recursive policy improves robot control with persistent memory
Tiny Recursive Policy: Recursive Latent Refinement for Visuomotor Control
Robotics
Summary
Robots often struggle to remember important details from earlier moments when deciding what to do next. The authors present Tiny Recursive Policy (TRP), a method that helps robots keep and update a memory of their past actions and surroundings over time. TRP refines this memory repeatedly within each step before deciding on the next moves. This approach works well on tasks needing long memory and complex decisions while using less computing power than other methods. It also adapts and stays robust even if the robot’s memory or visual input is disturbed.
What this means in practice
- •For robotics engineers: Design robot controllers that maintain and update memory efficiently to improve performance on complex, multi-stage manipulation tasks.
- •For automation system developers: Implement compact policies for industrial robots that can handle diverse tasks with fewer parameters and better robustness to unexpected inputs.
Authors
Hakan Aktas, Cengiz Oztireli
Abstract
Behaviour cloning has become a powerful approach for robotic manipulation, but many modern policies reconstruct their internal representation from a fixed observation window at each prediction and do not maintain persistent latent context throughout execution. We introduce Tiny Recursive Policy (TRP), a parameter-efficient behaviour-cloning policy that combines persistent temporal state with recursive latent refinement. At each control step, TRP repeatedly updates high- and low-level latent states using a shared computation module before decoding an action sequence, and propagates the resulting state to subsequent predictions. We evaluate TRP across LIBERO-100, MetaWorld ML45, and the LIBERO-Mem benchmarks. TRP performs competitively across multi-task and few-shot settings and shows substantial gains on long-horizon and context-dependent tasks, while using significantly fewer parameters than large generative baselines. Execution-time perturbations show that reliance on the persistent carry increases with temporal-context demand, while training with carry perturbations substantially improves robustness to both carry and visual corruptions. Ablations reveal a strong interaction between persistent state and recursive refinement: the benefit of carry depends strongly on the amount of within-step computation, and the resulting performance cannot be explained by parameter count alone.