PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Efficient Robotic Manipulation Models with Self-Recursive Pruning

Zijia Chen, Yuenan Hou, Yu Li, Weijie Li, Li Liu

arXiv:2610.08555v1cs.RO

TL;DR

Generic pruning criteria often target weight magnitude, reconstruction, or likelihood rather than closed-loop robotic action behavior. LCAM combines action-conditioned connection scoring with recursive coarse-to-fine pruning, and reports competitive performance across LIBERO suites and OpenVLA, including 84.0% success at 50% pruning on LIBERO-Object with OpenVLA.

  • Problem

    Generic pruning criteria target proxies such as weight magnitude, reconstruction, or likelihood, while robotic policies are ultimately evaluated by closed-loop task success.

    Source: §I · p. 1

  • Method

    LCAM ranks connections using row-normalized weight contribution, activation moments, and action-loss sensitivity, then recursively recalibrates pruning masks using offline demonstrations and distortion.

    Source: §I · p. 1, A. Loss-Conditioned Activation-Moment… · p. 3, B. Self-Recursive Coarse-to-Fine… · p. 4

  • Results

    Across three MATE suites LCAM retains substantial success at high sparsity, and on OpenVLA LIBERO-Object it achieves 84.0% success at 50% pruning.

    Source: §I · p. 2, C. Results · p. 5, §V · p. 6

  • Takeaways & Limitations

    The results support loss-conditioned, activation-aware importance estimation for pruning robotic policies without recovery training or simulator rollouts.

    Source: §I · p. 1, §V · p. 6

    Offline action distortion may not capture long-horizon interaction effects at extreme sparsity, and evaluation covers only three MATE suites and one OpenVLA setting.

    Source: §V · p. 7

Abstract

from arXiv · show

Network pruning can reduce parameter redundancy in robotic policies. However, generic pruning criteria are tailored for image recognition tasks and commonly designed to preserve weight magnitude, local reconstruction, or language-model likelihood rather than closed-loop action behavior. Directly applying these pruning algorithms to robotic tasks yields unsatisfactory performance. In this paper, we propose Loss-Conditioned Activation-Moment (LCAM) pruning, a training-free method for unstructured pruning of pre-trained robotic manipulation policies. Specifically, we first rank connections using row-normalized weight contribution, activation moments measured on calibration demonstrations, and the sensitivity of output directions to the action-prediction loss. We further design a self-recursive coarse-to-fine procedure: importance is recalibrated after each nested coarse pruning stage, while held-out offline action distortion guides fine-grained budget allocation after a sparsity knee. Our algorithm is free from costly recovery training and simulator rollouts after pruning. Experiments on three LIBERO suites with competitive robotic policies, together with evaluations on OpenVLA, show that LCAM attains competitive performance across a broad range of pruning ratios. Notably, on LIBERO-Object with OpenVLA, our LCAM achieves 84.0% success at 50% unstructured pruning, retaining over 90% of the dense policy's success rate. Promising results on real-world robotic ping pong further demonstrate the effectiveness of our pruning algorithm.

I. INTRODUCTION

Robotic policies are costly to run, and pruning them with generic proxy objectives may not preserve closed-loop action behavior. LCAM combines action-conditioned connection scoring with recursive pruning, achieving competitive results across robotic-policy benchmarks.

I. INTRODUCTION

  • High-capacity perception and policy networks create storage and inference costs, motivating pruning that reduces redundancy while preserving task performance.The paper focuses on unstructured pruning of individual weights in pretrained policies.Source: §I · p. 1
  • Robotic policies are evaluated by closed-loop task success, where small pruning-induced action deviations can accumulate; generic activation or reconstruction criteria may not capture this behavior.Robotic demonstrations are typically scarce and task-specific, and policy modules can differ in sensitivity to parameter removal.Source: §I · p. 1
  • LCAM ranks connections using row-normalized weight contribution, calibration-demonstration activation evidence, and action-loss sensitivity, then recursively recalibrates as sparsity increases.It uses offline action distortion to switch to finer pruning and requires neither post-pruning retraining nor simulator rollouts.Source: §I · p. 1
  • On LIBERO-Object with OpenVLA, LCAM achieves 84.0% success at 50% pruning and retains over 90% of the dense policy’s success rate.LCAM also achieves the highest success among compared baselines at 50% and 60% pruning.Source: §I · p. 2

A. Network Pruning

  • Magnitude pruning ranks parameters by absolute weight alone, while later methods use signals such as loss sensitivity, activations, reconstruction, or searched metrics.These criteria were primarily developed for image classification or language modeling, so their reliability for preserving robotic action behavior remains an empirical question.Source: A. Network Pruning · pp. 2–3

B. Model Compression and Pruning for Robotic Policies

Prior robotic-policy compression methods commonly pair structural reduction with retraining, distillation, or inference-time correction. LCAM instead constructs pruning masks without recovery optimization, using a connection-level criterion and validation-controlled coarse-to-fine allocation.

B. Model Compression and Pruning for Robotic Policies

  • LightDP, RLRC, and Shallow-π couple structural compression with retraining, fine-tuning, reinforcement learning, or distillation to preserve policy performance.GLUESTICK instead adds low-rank layer-wise corrections during inference after structural pruning.Source: B. Model Compression and Pruning for… · p. 3

III. METHOD

  • LCAM combines connection-level importance ranking with validation-controlled coarse-to-fine allocation to set a pruning budget without retraining or rollout-based mask selection.Section III-A defines the score, while Section III-B presents the allocation procedure.Source: §III · p. 3

A. Loss-Conditioned Activation-Moment Importance

LCAM estimates connection importance from the pretrained policy’s native action-prediction objective, combining relative weight contribution with activation evidence conditioned on action-loss sensitivity. Its calibration uses offline demonstrations, with a scale-aware cost for comparisons across layers or modules.

A. Loss-Conditioned Activation-Moment Importance

  • A first-order pruning-sensitivity approximation motivates retaining connections with non-negligible relative weight, informative activations, and influence on action-sensitive output directions.The action-prediction objective is evaluated on calibration demonstrations to obtain gradients; LCAM does not introduce a separate pruning loss or optimize that objective during mask construction.Source: A. Loss-Conditioned Activation-Moment… · p. 3
  • LCAM uses a two-pass calibration procedure to estimate layer-wise gradient scale and collect activation statistics weighted by action-loss sensitivity.Calibration demonstrations are fixed and separate from the held-out allocation set; the weighted moments retain sample-wise coupling between activation evidence and loss sensitivity.Source: A. Loss-Conditioned Activation-Moment… · p. 3
  • The score combines row-normalized weight contribution with joint activation moments, favoring connections repeatedly active in action-sensitive contexts and reducing output-row scale differences.Connections with lower scores are considered less important within their layer.Source: A. Loss-Conditioned Activation-Moment… · p. 3
  • For comparisons across layers or modules, LCAM rescales the row-normalized score by the output-row ℓ1 norm and layer-wise gradient scale.The pruning scope and retention constraints are fixed before mask construction.Source: A. Loss-Conditioned Activation-Moment… · p. 3

B. Self-Recursive Coarse-to-Fine Allocation

The procedure uses held-out offline action distortion to allocate pruning after a coarse-stage safe boundary, without rollouts or recovery training. LCAM is recalibrated as nested pruning proceeds, and fine-stage choices minimize distortion increases.

B. Self-Recursive Coarse-to-Fine Allocation

  • A disjoint held-out set measures normalized action deviation from the dense policy, while calibration and allocation use separate demonstration sets.For action-token policies, tokens are decoded into continuous robot actions before distortion is computed.Source: B. Self-Recursive Coarse-to-Fine… · p. 4
  • At each nested coarse stage, LCAM is recalibrated and marginal distortion is tracked to identify a safe boundary for fine allocation.Sparsity rises in fixed increments; the current candidate is rejected when the knee criterion is triggered.Source: B. Self-Recursive Coarse-to-Fine… · p. 4
  • From that boundary, the fine stage prunes in small increments, selecting the feasible group with the smallest action-distortion increase per removed connection.LCAM ranks connections within each candidate group; offline distortion determines which group is selected next.Source: B. Self-Recursive Coarse-to-Fine… · p. 4
  • Masks remain nested, and pruning uses no parameter optimization, recovery training, distillation, simulator rollouts, or test-success-guided selection.The method measures offline action distortion rather than task success when constructing masks.Source: B. Self-Recursive Coarse-to-Fine… · p. 4

A. Experimental Setup

The evaluation covers three LIBERO suites using MATE and an OpenVLA checkpoint fine-tuned on LIBERO-Object. Disjoint calibration and allocation demonstrations support pruning-mask selection without test trajectories or task success.

a) Benchmarks and Policies

  • MATE is the primary backbone across LIBERO-Spatial, LIBERO-Object, and LIBERO-Goal; OpenVLA-7B is additionally evaluated on LIBERO-Object.The OpenVLA checkpoint is fine-tuned on LIBERO-Object, with observation and evaluation protocols following the original model protocols.Source: A. Experimental Setup · pp. 4–5

b) Data protocol

  • Disjoint calibration and held-out allocation demonstration streams estimate LCAM importance and offline action distortion, respectively, without using test trajectories or task success to select masks or hyperparameters.The allocation set is used only for offline action-distortion measurement.Source: A. Experimental Setup · p. 5

c) Pruning scope and configuration

  • All methods use fixed eligible parameter scopes and are evaluated at pruning ratios from 10% to 60%; LCAM uses 16 calibration and 16 validation batches.Coarse and fine increments are 10% and 1%, respectively, with knee-sensitivity multiplier β = 2.0.Source: A. Experimental Setup · p. 5

B. Baselines

The comparison includes training-free pruning baselines spanning random, weight-based, activation-aware, loss-sensitive, and automatically discovered criteria. Across MATE suites and OpenVLA, LCAM retains performance more consistently at high sparsity, with comparable real-world ping-pong return rates to the dense model.

B. Baselines

  • The training-free baselines include random, magnitude, Wanda, Taylor, Pruner-Zero-style gradient saliency, and Auto-Prune; methods share eligible tensor scope and calibration budget, with no recovery training.The criteria cover random, weight-based, activation-aware, loss-sensitive, and automatically discovered pruning approaches.Source: B. Baselines · p. 5

C. Results

  • No compared baseline behaves consistently robustly across all three suites as sparsity increases, although Taylor and Pruner-Zero remain competitive on LIBERO-Object up to 40% pruning.Gradient-based baselines are more resilient in selected settings but remain suite-dependent.Source: C. Results · p. 5
  • Random pruning reaches near-zero success after 20% pruning, while magnitude pruning fails abruptly on two suites and Wanda and Auto-Prune deteriorate sharply at high sparsity.Weight-only and activation-based criteria are sensitive to policy suite and target pruning ratio.Source: C. Results · p. 5
  • At 60% pruning, LCAM retains 65.5%, 68.5%, and 55.0% success on LIBERO-Spatial, LIBERO-Object, and LIBERO-Goal, respectively, with more consistent high-sparsity retention across the MATE suites.On OpenVLA fine-tuned for LIBERO-Object, LCAM achieves 84.0% success at 50% pruning and 68.5% at 60%, the best rates among compared methods at those ratios; real-world ball return is 86% versus 88% for the dense model across 100 trials.Source: C. Results · p. 5

D. Ablation Study

The ablations compare the self-recursive coarse-to-fine pruning procedure with one-shot and coarse-only variants, then assess the complementary components of the LCAM connection score.

a) Effect of self-recursive coarse-to-fine pruning

  • At 60% pruning, full coarse-to-fine retains 68.5% success while both coarse-only variants collapse; nested pruning also beats one-shot pruning at 40% and 50%.The procedure ablation is on LIBERO-Object and uses the full LCAM score.Source: Table III · p. 6, D. Ablation Study · p. 6
  • Nested coarse pruning recalibrates LCAM on the currently sparse policy; the full method adds action-aware fine-grained allocation after the detected sparsity knee.The procedure compares one-shot pruning, nested coarse pruning, and the full LCAM procedure.Source: D. Ablation Study · p. 6

b) Effect of the LCAM importance criterion

  • Removing action-loss sensitivity or activation moments degrades performance, while removing row-normalized weight contribution leads to near-complete failure.The score ablation holds nested coarse pruning and the pruning configuration fixed while changing the connection score.Source: Table IV · p. 6, D. Ablation Study · p. 6

V. CONCLUSIONS

LCAM is a training-free post-training method for unstructured pruning of robotic manipulation policies, and its evaluations report competitive performance at high sparsity. The conclusion also identifies limits in the offline distortion surrogate and the breadth of evaluation.

V. CONCLUSIONS

  • Across three LIBERO suites with MATE policies, LCAM retains substantial success at high pruning ratios, while OpenVLA achieves the highest compared success at 50% and 60% pruning.At 50% pruning on OpenVLA, LCAM retains over 90% of the dense policy’s success rate.Source: §V · p. 6
  • LCAM uses loss-conditioned activation evidence to rank connections and constructs nested masks from offline demonstrations without recovery training or simulator rollouts.It is presented as a training-free post-training method for unstructured pruning.Source: §V · p. 6
  • Offline action distortion may not capture long-horizon interaction effects at extreme sparsity, and broader generalization across architectures, platforms, and real-world tasks remains unestablished.The current evaluation covers three MATE suites and one OpenVLA setting.Source: §V · p. 7

APPENDIX

The appendix separates calibration, allocation, and final evaluation data, then details the MATE pruning scope, loss gradients, and coarse-to-fine allocation implementation.

A. Pruning Data Protocol

  • Calibration demonstrations estimate LCAM statistics, disjoint allocation data select masks using offline action distortion, and final evaluation data are reserved for reporting performance.Final evaluation data are never accessed during mask construction; MATE uses native splits, while OpenVLA uses RLDS-formatted LIBERO data.Source: A. Pruning Data Protocol · p. 7

B. MATE Pruning Implementation

  • For MATE, pruning covers linear layers except the frozen track branch, including the action-prediction head, with shared row- and column-retention safeguards.The safeguards avoid fully disconnecting individual input or output features.Source: B. MATE Pruning Implementation · p. 7
  • MATE’s native action-prediction objective is mean-squared error, evaluated on calibration batches only to obtain LCAM gradients.LCAM statistics use 16 calibration batches; offline action distortion is evaluated on 16 held-out allocation batches.Source: B. MATE Pruning Implementation · p. 7
  • The controller uses 10% coarse and 1% fine sparsity increments, recalculates LCAM at coarse candidates and fine-round starts, and accepts fine blocks by distortion increase per pruned connection.Fine candidate blocks are formed across functional groups; mask construction uses neither recovery training nor simulator rollouts.Source: B. MATE Pruning Implementation · p. 7

C. OpenVLA Pruning Implementation

OpenVLA uses action-token loss gradients to calibrate LCAM scores for selected Llama projections, with memory-bounded accumulation and a validation-controlled coarse-to-fine pruning schedule. Masks are nested across stages, with no recovery training or simulator rollouts during pruning.

C. OpenVLA Pruning Implementation

  • LCAM uses mean negative log-likelihood over supervised action-token positions as the action loss for gradients, estimated from 16 calibration samples, each one RLDS transition.The visual encoder remains unchanged.Source: C. OpenVLA Pruning Implementation · p. 7
  • The selected Llama projection matrices contain approximately 6.4 billion weights, so LCAM accumulates moments in two passes, processing Transformer blocks in chunks of two.For calls with more than 2048 token positions, a deterministic subset of 2048 is used for moment accumulation.Source: C. OpenVLA Pruning Implementation · p. 7
  • OpenVLA uses the MATE validation-controlled schedule, with 10% coarse and 1% fine global sparsity increments and a knee test after two coarse marginal distortions at β = 2.0.Source: C. OpenVLA Pruning Implementation · p. 7
  • All layers in a pruning stage are scored against the same parent checkpoint, and their masks are applied together after chunk calibration finishes.Masks are nested because zeros persist across stages; pruning uses neither recovery training nor simulator rollouts.Source: C. OpenVLA Pruning Implementation · pp. 7–8
Loading 2610.08555v1…