Research map / Vision-language and multimodal / Robot action and navigation
ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ElegantVLA, a plug-in RL-trained scheduler that adaptively recomputes or reuses the vision-LLM backbone and action-head denoising steps of frozen VLA policies (GR00T, CogACT) to speed inference while preserving success.
Problem
VLA models are costly to run at every control step, limiting control frequency for real-time robotic manipulation; existing acceleration methods use fixed rules or optimize single components.
Approach
A lightweight Maskable PPO scheduler, conditioned on first-layer CKA representation similarity, gripper/end-effector motion speeds, and episode progress, jointly selects a five-level Vision-LLM compute mode and a three-level action-head denoising reuse mode per control step, trained in two stages (rule-guided teacher shaping, then task-efficiency reward).
Key results
- On CogACT, ElegantVLA reaches 77.59% success with 3.72x speedup on Visual Matching and 72.54% with 3.77x on Variant Aggregation, versus 74.80% and 61.30% for the CogACT baseline.
- On GR00T in SimplerEnv, success rises from 64.00% to 65.88% with up to 2.55x average FLOPs speedup.
- On real-world Franka tasks with GR00T, success improves from 61.67% to 65.00% and control frequency from 13.8 Hz to 26.3 Hz (2.18x computation reduction).
Surprising findings
- A hand-designed CKA threshold scheduler improves Google Robot average but drops WidowX success from 57.93% to 54.71%, showing fixed rules do not transfer across regimes.
- Stage1 RL (rule-guided teacher) can be overly aggressive, e.g., reaching 2.43x speedup on MoveNear but dropping success to 80.00%.
Limitations
- Speedups reported mainly as FLOPs; wall-clock gains smaller (about 1.91x real robot)
- Gains on MoveNear are smaller, with the scheduler conservative for continuous spatial reasoning
- Real-world evaluation limited to 10 trials per task on one robot
Experimental setup
Datasets: SimplerEnv
Benchmarks: SimplerEnv Google Robot SimplerEnv WidowX SimplerEnv Visual Matching SimplerEnv Variant Aggregation Franka Research 3 real-world six tasks
Metrics: success rate FLOPs speedup latency control frequency
Baselines: CogACT GR00T Random Dropping FastV VLA-Cache MoLe-VLA
Builds on: GR00T CogACT FastV VLA-Cache PPO
Closest papers by meaning
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.94)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.93)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.93)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.93)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.93)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.