Research map / Vision-language and multimodal / Robot action and navigation
Vision-language-action foundation models
Trains and benchmarks end-to-end VLA policies, with reasoning, unified representations and efficient token caching.
264 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2024–2026) · an approach in Robot action and navigation, keyword cluster “vla · language action · robotic”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
7 papers in 2023–24 and 257 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 9.2% to 51.7%.
Open questions the data points to
- Papers report OpenVLA fine-tuned on LIBERO-Long (Success rate) anywhere from 45.7 to 53.7: comparisons against it depend on whose number you take.
- Papers report OpenVLA fine-tuned on LIBERO-Goal (Success rate) anywhere from 70.4 to 79.2: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2024 · Vision-Language Foundation Models as Effective Robot Imitators (ICLR 2024 · 19 citations)
- 2025 · CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · 58 citations)
- 2026 · OpenDriveVLA: Towards End-to-end Autonomous Driving with Large Vision Language Action Model (AAAI 2026 · 26 citations)
Earliest papers here
- Vision-Language Foundation Models as Effective Robot Imitators (ICLR 2024 · 19 citations)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · 14 citations)
- DeeR-VLA: Dynamic Inference of Multimodal Large Language Models for Efficient Robot Execution (NeurIPS 2024 · 14 citations)
What has been tried
45 papers here report 534 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 7%, hurt 22%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add visual chain-of-thought subgoal image generation | LIBERO-Spatial and LIBERO-Goal · Success rate | Helped | CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · 58 citations) | |
| Use disentangled path and speed waypoints instead of entangled waypoints | CARLA Leaderboard 2.0 · Driving Score | +39.9 percent | Helped | SimLingo: Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment (CVPR 2025 · 28 citations) |
| freeze the fusion decoder embedding layer | CALVIN ABCD→D (Enriched) · Avg Len | 1.85 -> 2.12 | Helped | Vision-Language Foundation Models as Effective Robot Imitators (ICLR 2024 · 19 citations) |
| use bi-level search instead of direct search | OVMM · Overall SR | 48.5% -> 49.4% | Helped | MoManipVLA: Transferring Vision-language-action Models for General Mobile Manipulation (CVPR 2025 · 16 citations) |
| Add CoT reasoning supervision instead of action-only supervision | nuScenes validation · L2 distance | 0.41 -> 0.40 (ST-P3 Avg.); 0.71 -> 0.70 (UniAD Avg.) | Mixed | AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · 16 citations) |
| Add closed-loop planning with RRC | Rigid object manipulation tasks · success rate | 51.7% -> 68.3% | Helped | OmniManip: Towards General Robotic Manipulation via Object-Centric Interaction Primitives as Spatial Constraints (CVPR 2025 · 15 citations) |
| Add predicted bounding boxes to the input prompt | Open-X test sets · PSNR | 17.02 -> 17.21 | Helped | 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · 14 citations) |
| add auxiliary losses at intermediate exits | CALVIN ABCD→D, 4.9 GFLOPs · average successful length | 2.64 -> 3.94 | Helped | DeeR-VLA: Dynamic Inference of Multimodal Large Language Models for Efficient Robot Execution (NeurIPS 2024 · 14 citations) |
| add VQA pre-training | nuScenes · L2 Avg. (m), UniAD metrics | 1.22 -> 1.19 | Helped | FutureSightDrive: Thinking Visually with Spatio-Temporal CoT for Autonomous Driving (NeurIPS 2025 · 13 citations) |
| Use Mamba-2.7B instead of RWKV-3B | General and robotic-related reasoning benchmarks · reasoning performance | Helped | RoboMamba: Efficient Vision-Language-Action Model for Robotic Reasoning and Manipulation (NeurIPS 2024 · 13 citations) | |
| Reuse visually static tokens without filtering task-relevant tokens | LIBERO Spatial · success rate | 84.4 -> 74.2 | Hurt | VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · 9 citations) |
| Fine-tune on Impromptu VLA before nuScenes | NeuroNCAP · average NeuroNCAP score | 1.77 -> 2.15 | Helped | Impromptu VLA: Open Weights and Open Data for Driving Vision-Language-Action Models (NeurIPS 2025 · 8 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| OpenVLA fine-tuned | LIBERO-Long | Success rate | 45.7 – 53.7 | 2 papers, e.g. ICLR 2025 and CVPR 2025 |
| OpenVLA fine-tuned | LIBERO-Goal | Success rate | 70.4 – 79.2 | 2 papers, e.g. ICLR 2025 and CVPR 2025 |
| OpenVLA fine-tuned | LIBERO-Object | Success rate | 83.8 – 88.4 | 2 papers, e.g. ICLR 2025 and CVPR 2025 |
What reviewers will expect
Datasets and baselines named in the 103 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- LIBERO · 25 of 103 papers
- Open X-Embodiment · 10 of 103 papers
- DROID · 8 of 103 papers
- nuScenes · 6 of 103 papers
- LIBERO-Plus · 6 of 103 papers
- CALVIN · 5 of 103 papers
- NAVSIM · 5 of 103 papers
- RoboTwin 2.0 · 5 of 103 papers
Baselines
- OpenVLA · 24 of 103 papers
- pi0 · 17 of 103 papers
- OpenVLA-OFT · 15 of 103 papers
- Diffusion Policy · 12 of 103 papers
- Octo · 10 of 103 papers
- pi0.5 · 9 of 103 papers
- UniVLA · 8 of 103 papers
- UniAD · 7 of 103 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 264 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 23 papers
- Shanghai Jiao Tong University · 13 papers
- Peking University · 12 papers
- Hong Kong University of Science and Technology · 9 papers
- Xi’an Jiaotong University · 8 papers
- The Chinese University of Hong Kong · 7 papers
- Fudan University · 5 papers
- Harbin Institute of Technology · 5 papers
- Alibaba Group · 5 papers
- Sun Yat-sen University · 5 papers
Use it
Download BibTeX of 20 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Memory and subgoal-based manipulation (103 papers)
- Embodied multimodal LLM agents (70 papers)
- Multimodal prompt-driven manipulation (50 papers)
Elsewhere in the topic
- Pretraining and data augmentation for VLN (97 papers)
- Reasoning-enhanced vision-language navigation (46 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).