Kurate.org
Rankings
Back to Rankings

SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models

Shenli Zheng, Zhou Fang, Yuan Du, Yijiang Liu, Li Du

Oct 19, 2026arXiv:2610.26487v1
cs.ROcs.AI
Share
Scorecard· 16/16
5.0/10 impact

Useful, timely applied study of VLA compression + RL recovery with real-hardware latency, but incremental over PAT/RLRC, lacking key baselines, variance reporting, and real-robot validation of the headline configuration.

Abstract

The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate. Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning alone recovers. We present SPAR (Structural Pruning and Action Recovery), which simultaneously prunes the hidden and feed-forward intermediate dimensions of the large language model (LLM) backbone through pruning-aware tuning, then applies reinforcement learning (RL) to recover policy performance. Applied to π0.5π_{0.5}, SPAR reduces the language backbone from 2.34B to 0.09B parameters and the full model from 3.37B to 1.12B, removing approximately 96% and 67% of their parameters, respectively. At this compression level, RL raises average success across the four LIBERO suites from 84.55% to 93.45%, an 8.90-percentage-point gain. The compressed policy achieves a 1.98×1.98\times full-policy inference speedup over the unpruned model. These results identify substantial intermediate-dimension redundancy and show that post-pruning RL can recover manipulation performance in compact policies.

AI Impact Assessments

(1 model)

Core Contribution

SPAR compresses the LLM backbone of the π0.5 VLA model by learning masks for two dimensions with separate retention targets: the hidden dimension and the FFN intermediate dimension. It builds on the existing PAT (pruning-aware tuning) method. After physical pruning and SFT, it applies PPO-based RL to recover manipulation performance. The action expert and value head are trained while the pruned backbone stays frozen. The headline result is that the LLM shrinks from 2.34B to 0.09B parameters (the full model from 3.37B to 1.12B). Average LIBERO success is 84.55% after PAT and 93.45% after RL, with a 1.98× latency speedup on Jetson AGX Orin.

The novelty is mostly in combination and engineering. The pruning stage is an adaptation of PAT with separate retention targets, sigmoid annealing, and a soft-count penalty. RL recovery after compression is already explored in RLRC, which the paper cites. The most interesting empirical finding is that intermediate-dimension pruning is nearly free, with 90% of FFN width removed at a cost of about 0.35 points on LIBERO. Hidden-dimension pruning is much more damaging.

Methodological Rigor

*Strengths:*

  • The paper sweeps retention targets systematically and reports latency and memory on real edge hardware (Jetson with TensorRT).
  • It includes a LIBERO-Plus robustness analysis and a small real-robot study.
  • It honestly reports a persistent robustness gap: 56.8% vs 84.5% for the smallest model.

    • *Weaknesses:*

  • The comparison with the unpruned baseline is confounded. LIBERO is near-saturated (96.8% baseline), so differences are small and probably within noise, yet no seeds, confidence intervals, or variance are reported. Real-robot results use only 20 trials per task.
  • The pruned-model success rates are single-run.
  • There is no comparison against other structured pruning methods applied to π0.5 (LLM-Pruner, SliceGPT, or naive width-reduced training from scratch). The only direct pruning comparison is a non-same-platform EcoVLA number.
  • RLRC, the closest prior work, is not compared experimentally.
  • The unpruned model also receives RL, but the RL-versus-SFT-budget comparison is not controlled. PAT-only models get 30k SFT steps, while RL adds substantial extra environment interaction on the very LIBERO tasks used for evaluation. RL gains therefore partly reflect extra training on the test distribution (in-simulator), not only "recovery".
  • The hidden and intermediate masks are shared across all layers, which is a restrictive design. Rankings and ablations are narrow: the loss ablation differences (0.35–0.75 points) are within plausible noise.
  • Because the backbone is frozen during RL, recovery comes only from the action expert and value head. This is a notable design finding, but it is not ablated against full fine-tuning.
  • Real-robot evaluation covers only intermediate-only pruning (20%), not the headline SPAR configuration. This is a significant gap, since the headline claim (0.09B LLM) is never validated on hardware.
  • The LIBERO-Plus results show that the aggressively compressed model is much less robust, which tempers the practical claims.

    • Potential Impact

    The work addresses a real deployment bottleneck: running VLAs on onboard compute. The practical takeaway (prune FFN width aggressively, be careful with hidden-dimension pruning, and use RL to recover) is actionable. A 1.98× speedup on Jetson is meaningful, but the latency remains about 149 ms for the full policy. The flow-matching action expert, vision encoder, and other components are unchanged, which limits speedup (Amdahl's law). The 1.12B model with 93.45% success is still well below the baseline.

    The impact is likely to be moderate and incremental, within the VLA-efficiency subfield. The method is tied to PAT and π0.5, and the evaluation is limited to LIBERO. The paper is unlikely to become a foundational reference, but it provides useful empirical data points on how redundant Gemma-2B is for manipulation.

    Timeliness

    VLA efficiency and RL post-training of VLAs are both active topics (SimpleVLA-RL, RLinf, π*0.6). The combination of the two is timely, though several concurrent works (RLRC, EcoVLA) occupy the same space, which reduces novelty and priority.

    Strengths

  • Clear, practical finding on dimension sensitivity.
  • Real hardware latency and memory measurements.
  • Robustness analysis that honestly exposes a limitation.
  • Clean description of mask learning with convergence plots.

    • Limitations

  • No variance or statistical testing.
  • Missing baselines such as other structured pruning methods and RLRC.
  • One backbone, one benchmark family, and the real-robot test does not cover the headline configuration.
  • Incremental over PAT and RLRC.
  • Mostly simulation evidence from the same tasks used for RL training.
  • No code release mentioned. Details are reasonably specified (batch sizes, steps, PPO configuration), but the λ values and many RL hyperparameters are missing.
  • Resource requirements are moderate to high: 8×A800 for RL.

    • Overall this is a competent, applied, incremental contribution with useful empirical observations but limited conceptual novelty and incomplete baselines.

    Rating:5/ 10
    Significance 5Rigor 4.5Novelty 4Clarity 7

    Generated Oct 20, 2026

    Comparison History (0)

    No comparisons yet.