Research map / Vision-language and multimodal / Robot action and navigation
LEAP-VLA: Latent-Enhanced Action Prototyping via Continuous Residual Latent Spaces for Vision-Language-Action Models
Bo-Yun Yu, Kuan-Chuan Peng, Jun Hsieh
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 |
|---|---|---|---|
| predict latent zq targets instead of raw actions (ablation) | LEAP-VLA aligner with linear head · LIBERO-Spatial | success rate (%): 92.4 -> 94.6 (+2.2%) | Improved |
| replace linear head with Prototype-Aware Output Head (ablation) | LEAP-VLA aligner · LIBERO-Spatial | success rate (%): 94.6 -> 98.4 (+3.8%) | Improved |
| use multi-level residual quantization instead of single-level soft assignment (ablation) | MSRQ · LIBERO-Spatial | MAE (x10^-3): 2.5 -> 1.5 | Improved |
| use L2-based soft assignment instead of discrete hard assignment (MSRQ-single vs discrete multi-level) (ablation) | MSRQ-single · LIBERO-Spatial | MAE (x10^-3): 2.5 vs 7.7 (SimRVQ), 9.0 (RVQ-Commit) | Improved |
| use soft residual levels instead of discrete residual levels for utilization (ablation) | MSRQ vs RVQ-Commit · LIBERO-Spatial level-0 utilization | entropy ratio (%): 51.9% -> 99.5% | Improved |
| use bidirectional Transformer with CNN (BiCT) instead of causal Transformer (ablation) | MSRQ encoder/decoder · LIBERO-Spatial | MAE (x10^-3): 71.9 -> 1.5 | Improved |
| use BiCT instead of CNN only (ablation) | MSRQ encoder/decoder · LIBERO-Spatial | MAE (x10^-3): 8.5 -> 1.5 | Improved |
| use BiCT instead of MLP only (ablation) | MSRQ encoder/decoder · LIBERO-Spatial | MAE (x10^-3): 15.4 -> 1.5 | Improved |
| use 1B LEAP-VLA instead of larger robot-pretrained VLAs | LEAP-VLA · LIBERO average | success rate (%): 92.7 (UD-VLA 8.5B) -> 97.1 | Improved |
| use LEAP-VLA without robot pretraining vs OpenVLA-OFT with robot pretraining | LEAP-VLA · CALVIN ABC->D | average sequence length: 4.10 -> 4.28 | Improved |
| use LEAP-VLA instead of VLM4VLA (VLM-only) | LEAP-VLA · CALVIN ABC->D | average sequence length: 4.06 -> 4.28 | Improved |
| use LEAP-VLA instead of pi0 on LIBERO-Long | LEAP-VLA · LIBERO-Long | success rate (%): 85.2 -> 93.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- Fast-ThinkAct — Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning (CVPR 2026) · better on 1 of 1 shared result
- UP-VLA — UP-VLA: A Unified Understanding and Prediction Model for Embodied Agent (ICML 2025) · better on 1 of 1 shared result
- VLAS — VLAS: Vision-Language-Action Model with Speech Instructions for Customized Robot Manipulation (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.92)
- RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.92)
- ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.92)
- CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification (NeurIPS 2025 · similarity 0.91)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
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.
