Efficient Vision-Language-Action Models for Embodied Manipulation: A Systematic Survey

Technical Summary: Efficient Vision-Language-Action Models for Embodied Manipulation: A Systematic Survey

Problem Statement
Vision-Language-Action (VLA) models have emerged as a dominant paradigm for embodied control, mapping natural language instructions and visual observations directly to robot actions. While these models offer superior semantic understanding and generalization compared to traditional task-specific algorithms, they face significant efficiency challenges. Contemporary VLA systems often rely on large pre-trained vision-language models (VLMs) and heavy visual backbones, resulting in massive parameter counts, high memory footprints, and slow inference speeds. These characteristics conflict with the strict constraints of edge platforms, such as on-board mobile manipulators, which require real-time performance, limited computational resources, and low energy consumption. Although efficiency optimization has been extensively studied in VLMs, directly applying those techniques to VLAs is non-trivial due to the additional requirements for temporally consistent action sequences, real-time constraints, and physical reliability. This survey addresses the gap in existing literature by providing the first systematic review dedicated specifically to the efficiency of VLA models.

Methodology and Taxonomy
The authors categorize existing solutions for improving VLA efficiency into four primary dimensions, following the processing pipeline of VLA systems:

1. Model Architecture: Strategies for designing computationally efficient foundational structures.
2. Perception Feature: Methods for reducing front-end computational costs related to visual inputs.
3. Action Generation: Techniques for accelerating the production of control signals.
4. Training/Inference Strategies: Optimization techniques across the model lifecycle, including training paradigms and deployment inference.

Key Contributions and Technical Details

• Efficient Model Architectures:

  • Static Backbone Selection: The survey reviews approaches that replace large-scale backbones with lightweight alternatives. Examples include using state-space models like Mamba (RoboMamba) for efficient temporal modeling, or employing smaller language models (e.g., Pythia-1.3B in TinyVLA, SmolVLM in SmolVLA) to reduce parameter counts while maintaining core task capabilities.
  • Dynamic Computation Pathways: To preserve the expressive power of large models while reducing inference cost, methods introduce dynamic routing. This includes layer pruning (FLOWER), early-exit mechanisms with lightweight policy heads (DEER-VLA), Mixture-of-Experts frameworks that dynamically select active layers (MoLE-VLA), and similarity-based layer skipping (Efficient-VLA).
  • Dual-System Design: Inspired by cognitive science, this approach divides the model into a "slow" system for complex reasoning (using large MMLMs) and a "fast" system for rapid, intuitive responses (using lightweight models). Examples include LCB, HiRT, RoboDual, and OpenHelix, which use latent tokens or special tokens to communicate between systems. Non-standard variants like HyperVLA use a HyperNetwork to generate parameters for a compact Base Policy.

• Efficient Perception Features:

  • Selective Feature Processing: Since visual inputs dominate token sequence length, methods focus on pruning or compressing redundant visual tokens. Techniques include attention-based token pruning (FastV, EfficientVLA), spatial-aware pruning using edge detection (SP-VLA), and feature-based scoring via Singular Value Decomposition (FlashVLA). Other approaches use query-driven selection (LightVLA) or action-aware dynamic pruning (ADP, FASTDriveVLA).
  • Temporal Sharing and Reuse: To exploit temporal redundancy in sequential decision-making, methods reuse computation results across time steps. This includes reusing Key-Value (KV) caches for static image patches (VLA-cache), fusing visual tokens across frames based on importance masks (TTF-VLA), and caching high-level reasoning outputs or diffusion intermediate features (Fast ECoT, Efficient-VLA).

• Efficient Action Generation:

  • Raw Action Generation: To mitigate error accumulation and improve throughput, methods generate action chunks rather than single steps. Techniques include temporal ensembling (Action Chunk), sequence inpainting for smooth transitions (RTC), and compressing action sequences via Discrete Cosine Transform or Byte-Pair Encoding (FAST, OmniSAT). Some models use a single special token to represent an entire action chunk (VOTE).
  • Reasoning-Aware Action Generation: To improve generalization and interpretability, some models incorporate explicit reasoning steps. Language-based reasoning decomposes tasks into subtasks (ECoT), with efficiency gains achieved through plan reuse (Fast ECoT) or reasoning dropout at test time (ECoT-Lite). Vision-based reasoning extracts spatial-semantic cues, such as predicting goal-state frames (UniPi, SuSIE) or dynamic image regions (DreamVLA), often leveraging pre-trained video diffusion models.

• Efficient Training and Inference:

  • Training Efficiency: The survey covers Parameter-Efficient Fine-Tuning (PEFT) like LoRA, knowledge distillation (CEED-VLA, MoLE-VLA, VITA-VLA) to transfer capabilities from large to small models, structured parameter pruning with recovery mechanisms (GLUESTICK), and quantization-aware training (SQIL, BitVLA) to enable low-bit inference.
  • Inference Efficiency: To overcome the latency of autoregressive decoding, methods explore non-autoregressive or parallel decoding. This includes bidirectional attention for parallel sequence prediction (openvla-oft), speculative decoding with a draft model (Spec-VLA), and iterative refinement via Jacobi iteration with consistency distillation (PD-VLA, CEED-VLA).

Results and Analysis
The survey synthesizes representative works within each category, analyzing their strengths and weaknesses. It highlights that while static backbones offer direct efficiency gains, they may compromise generalization. Dynamic pathways offer flexibility but require careful design of selection criteria. Dual-system architectures balance reasoning and speed but can introduce latency due to asynchronous operation. In perception, token pruning reduces compute but risks losing critical information if thresholds are not adaptive. In action generation, raw action methods are faster but lack interpretability, while reasoning-based methods improve generalization at the cost of increased latency.

Significance and Future Prospects
The paper claims to be the first systematic survey dedicated to efficient VLAs, filling a critical gap as efficiency becomes a central bottleneck for real-world adoption. It argues that future research should move beyond simple model compression toward:

1. Co-optimization of Models and Data: Addressing data inefficiency and the "data pyramid" to balance model capacity with data quality.
2. Efficient Spatio-Temporal Perception: Developing compact 3D representations and temporal memory structures that avoid token explosion.
3. Compact Continuous Control: Shifting from chunked commands to continuous, context-aware control with hierarchical abstractions.
4. Resource-Aware Reinforcement Adaptation: Combining imitation learning with efficient reinforcement learning strategies to improve autonomy without prohibitive sample costs.
5. Efficiency-Centric Benchmarking: Establishing standardized evaluation frameworks that report resource efficiency, task performance, and interpretability together.

The authors conclude that balancing capability and efficiency is essential for the development of practical, generalizable, and resource-conscious embodied intelligence.