PaperlayerView exact revision on arXiv

Source-linked AI summary

Tiny Recursive Policy: Recursive Latent Refinement for Visuomotor Control

Hakan Aktas, Cengiz Oztireli

arXiv:2610.33288v1cs.RO

TL;DR

Many robotic policies reconstruct context from a fixed observation window, although interaction history can matter when visually similar observations call for different actions. TRP combines persistent latent state with recursive refinement. Across the evaluated benchmarks, it performs competitively in multi-task and few-shot settings and shows its strongest advantages on long-horizon and context-dependent tasks.

  • Problem

    Many policies reconstruct internal context from a fixed observation window, although interaction history can distinguish visually similar observations that require different actions.

    Source: §1 · p. 1

  • Method

    TRP repeatedly refines high- and low-level latent states with a shared module and propagates the resulting state across policy steps.

    Source: §1 · p. 2

  • Results

    Across LIBERO, MetaWorld ML45, and LIBERO-Mem, TRP performs competitively in multi-task and few-shot settings and shows its strongest advantages on long-horizon and context-dependent tasks.

    Source: §5 · p. 9

  • Takeaways & Limitations

    Carry reliance varies with task: removing it has almost no effect on MetaWorld but substantially degrades performance on long-horizon and context-dependent tasks.

    Source: §5 · p. 9

    Evaluation is restricted to simulation, so real-robot experiments are needed to establish whether the benefits persist in real-world settings.

    Source: §5 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

Behaviour-cloning policies have advanced through action chunking, latent actions, and generative methods, but many still reconstruct context from a fixed observation window. TRP combines persistent temporal state with recursive refinement for parameter-efficient visuomotor control.

1 INTRODUCTION

  • Action chunking predicts coherent future action sequences, latent-action methods learn reusable representations of extended behaviour, and generative policies use iterative action generation.The cited examples include Zhao et al. (2023), Lee et al. (2024), Mete et al. (2024), Chi et al. (2025), Reuss et al. (2023), and Pan et al. (2026).Source: §1 · p. 1
  • In long-horizon or partially observable tasks, interaction history can help distinguish visually similar observations that require different actions.Many policies instead reconstruct their internal state from a fixed temporal window at each query.Source: §1 · p. 1
  • Recursive refinement lets later computation revise intermediate representations conditioned on the current observation and temporal state, increasing expressivity without increasing model size.Source: §1 · pp. 1–2
  • TRP combines persistent temporal state and recursive latent refinement, repeatedly updating high- and low-level states before generating actions and carrying them into subsequent predictions.It also introduces training-time carry perturbations to improve robustness.Source: §1 · p. 2

2 RELATED WORK

Related work covers behaviour-cloning approaches for multimodal actions and temporal structure, generative policies, and recursive computation. TRP draws on these strands while applying recurrent latent refinement to robotic control.

Behaviour Cloning

  • Behaviour-cloning approaches address multimodal actions with mixture-density, energy-based, and autoregressive policies, while chunking and latent skills represent temporally extended behaviour.Examples include ACT (Zhao et al., 2023) and QueST.Source: §2 · p. 2

Generative policies

  • Diffusion and flow policies use iterative generation for continuous control; MIP (Pan et al., 2026) matches flow-based policies with two regression-based refinement steps.Source: §2 · p. 2

Recursive Reasoning

  • HRM (Wang et al., 2025) uses high- and low-level recurrent modules at different timescales, while TRM (Jolicoeur-Martineau, 2025) showed much of this complexity was unnecessary, replacing the two-module hierarchy.Source: §2 · p. 2
  • Recursive refinement has been applied to control and image generation, but nested within-token refinement in Autoregressive TRMs (Rauba et al., 2026) showed no advantage over standard recurrent Transformers.The passage suggests benefits may depend on repeatedly revising a candidate solution rather than committing autoregressively to individual outputs.Source: §2 · p. 3

3 METHOD

TRP combines recursive refinement of high- and low-level latent states with a persistent carry reused across control steps. It decodes future action sequences from the refined state, while training detaches carry between policy evaluations and can add carry perturbations for robustness.

3 METHOD

  • TRP recursively updates high- and low-level latent states and propagates the resulting pair as a persistent carry across control steps.The carry is reused at the next timestep, while gradients are detached between consecutive policy evaluations during training.Source: §3 · pp. 3–4
  • The observation window is encoded into a conditioning representation and converted into action-query embeddings for the recurrent computation core.The core operates over future action slots rather than directly over the raw observation sequence.Source: §3 · p. 3
  • Each of R high-level cycles contains K low-level refinement steps, with the low-level state refined before the high-level update using a shared module.The same computation module is reused for both update types.Source: §3 · pp. 3–4
  • After refinement, the action head decodes a future action sequence from the high-level state, which is generally preferred over alternative latent combinations.The resulting carry is passed to the next policy iteration.Source: §3 · p. 4
  • Training can add an auxiliary prediction loss from a perturbed carry to improve stability when persistent latent states are noisy or imperfect.This perturbation objective does not change the inference-time policy.Source: §3 · pp. 4–5

4 EXPERIMENTAL RESULTS

Across multi-task, few-shot, and context-dependent evaluations, TRP is strongest on LIBERO-Mem, while its advantage is smaller on standard benchmarks. Ablations show that carry benefits depend on task context and sufficient recursive refinement, and are not explained by model size alone.

4.1 MULTI-TASK LEARNING

  • TRP and TRP-perturbed outperform all baselines on LIBERO-90, and their advantage over the strongest non-TRP baseline widens on LIBERO-LONG.On MetaWorld ML45, several methods cluster near the top; TRP-perturbed has the highest mean by a small margin, while TRP-no-carry performs substantially worse.Source: §4.1 · p. 6, Figure 1 · p. 6

4.2 FEW-SHOT GENERALISATION

  • 0.704 success makes TRP the strongest model on five-shot LIBERO-LONG, with TRP-perturbed performing similarly; ACT leads MetaWorld few-shot, while TRP-no-carry falls short of every policy on LIBERO-LONG.Removing carry reduces LIBERO-LONG performance to roughly half that of the full model, a larger degradation than in standard multi-task training.Source: §4.2 · p. 6

4.3 EVALUATION ON CONTEXT-DEPENDENT TASKS

  • On LIBERO-Mem, standard TRP achieves a subgoal-completion score of 0.597, while the strongest non-TRP methods remain around 0.41; this setting produces the largest separation between TRP and comparison methods.Source: §4.3 · p. 7
  • Removing carry substantially reduces LIBERO-Mem performance, while carry perturbation yields an intermediate result, consistent with persistent state carrying useful task context across evaluations.Diffusion Policy and QueST, though competitive in other settings, remain substantially below TRP on LIBERO-Mem.Source: Figure 1 · p. 6, §4.3 · p. 7

4.4 ABLATIONS

  • Carry use at execution varies by benchmark: wiping it reduces TRP by approximately 37% on LIBERO-LONG and 42% on LIBERO-Mem, but has almost no effect on MetaWorld ML45.On MetaWorld, TRP falls only from 0.943 to 0.941 when carry is wiped, whereas TRP-no-carry reaches 0.742; the passage attributes that gap to training with carry rather than online temporal context.Source: Table 2 · p. 7, §4.4 · p. 8
  • Carry perturbation training improves robustness to carry removal and visual blackout on LIBERO-LONG and LIBERO-Mem; under long blackout, TRP-perturbed retains 0.674 LIBERO-LONG success versus 0.438 for TRP.The gains come with little cost to clean performance, although standard TRP remains stronger on LIBERO-Mem.Source: Table 2 · p. 7, §4.4 · p. 8
  • Carry and recursion interact: with shallow H1L1 recursion, adding carry lowers LIBERO-LONG success from 0.642 to 0.206, while carry remains mildly beneficial on LIBERO-Mem.With the default recursive configuration, enabling carry substantially improves performance across all three benchmarks; the authors suggest shallow models may not adequately reconcile carry with current observations.Source: Table 3 · p. 7, §4.4 · p. 9
  • Propagating gradients through the carry reduces performance on LIBERO-LONG and LIBERO-Mem, with longer truncated backpropagation horizons causing further degradation.TRP’s default carries forward state while detaching it between timesteps.Source: Table 3 · p. 7, §4.4 · p. 9

5 CONCLUSION

TRP combines persistent temporal context with recursive latent refinement and performs competitively across multi-task and few-shot settings, with its strongest advantages on long-horizon and context-dependent tasks. The evaluation is limited to simulation, and the learned carry’s contents and appropriate computation depth remain open questions.

5 CONCLUSION

  • TRP performs competitively across multi-task and few-shot settings, with its strongest advantages on long-horizon and context-dependent manipulation tasks.The conclusion presents persistent state paired with sufficient recursive refinement as a parameter-efficient mechanism for temporally demanding control.Source: §5 · pp. 9–10
  • The evaluation is restricted to simulation, so real-robot experiments are needed to establish whether the benefits persist in real-world settings.Source: §5 · p. 9
  • Interventions identify when TRP relies on its carry but do not reveal what it encodes or how long information is retained; fixed recursion is beneficial only up to a point.The conclusion motivates probing the carry and adaptive computation that allocates refinement to the current decision.Source: §5 · p. 10

Related papers

Loading 2610.33288v1…