Research map / Architectures and efficiency / RNNs and normalization
Characterizing Vision-Language-Action Models across XPUs: Constraints and Acceleration for On-Robot Deployment
Kaijun Zhou, Qiwei Chen, Da Peng, Zhiyang Li, Xijun Li, Jinyu Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compile π0 with torch.compile / ATC | π0 · RTX 4090 | latency (ms): 102.3 -> 35.2 | Improved |
| compile π0 with torch.compile | π0 · Jetson Thor | latency (ms): 246.0 -> 163.0 | Improved |
| add DP-Cache (S=4) to diffusion policy | Diffusion Policy · LIBERO Spatial/Object/Long, RTX 4090 | latency (ms): 378 -> 200 | Improved |
| increase DP-Cache cache steps from S=4 to S=8 (ablation) | DP-Cache · LIBERO Spatial | success rate (%): 78.9 -> 71.6 | Hurt |
| add DP-Cache (S=4) to π0.5-droid on Franka | π0.5-droid · Franka table cleaning (50 trials) | success rate (%): 76 -> 70 | Hurt |
| add ATC compilation to DP-Cache on Ascend 310P (ablation) | DP-Cache · Diffusion Policy, Ascend 310P | speedup: 3.5× -> 6.0× | Improved |
| add 4-bit quantization to Spec-VLA (ablation) | OpenVLA · LIBERO average | success rate (%): 75.1 -> 68.5 (Spec. vs Spec.+Cache); Quant. alone 76.5 -> 67.7 | Hurt |
| combine Spec-VLA with Cache-VLA (ablation) | OpenVLA · LIBERO average | speedup: 1.11× -> 1.29× speedup; SR 75.1 -> 68.5 | Mixed |
| increase stale denoising steps in V-AEFusion from 5 to 7 (ablation) | V-AEFusion · JAKA S5 Putting Block in Bowl | success rate (%): 88 -> 50 | Hurt |
| use 1 to 3 stale denoising steps in V-AEFusion (ablation) | V-AEFusion · JAKA S5 Putting Block in Bowl | success rate (%): 92 -> 90 | No clear effect |
| add V-AEFusion pipeline to π0 | π0 · RTX 4090 | latency (ms): 102 -> 77 | Improved |
| add V-AEFusion pipeline to π0 | π0 · Ascend 310P | latency (ms): 818 -> 820 | No clear effect |
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.90)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.88)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.88)
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.
