Research labChina
Open research collective (authors from Westlake University, BUPT, HKUST-Guangzhou and others) that released VLA-Adapter, ReconVLA and Spatial Forcing.
OpenHelix (GitHub organisation 'OpenHelix Robotics', created June 2025) is a research collective whose papers list authors from Westlake University, Beijing University of Posts and Telecommunications, Zhejiang University and HKUST (Guangzhou). It released VLA-Adapter (September 2025), a VLA with a 0.5B Qwen2.5 backbone and a 97M-parameter bridge-attention policy, as well as ReconVLA and Spatial Forcing, with weights on Hugging Face.
Sources: github.com (19 Oct 2026) · arxiv.org (11 Sep 2025) · arxiv.org (14 Aug 2025) · arxiv.org (14 Oct 2025)
Research lab · Academic research collective; no funding disclosed
#6 of 104 model companies · score 85.2 / 100 · updated 19 Oct 2026
Each bar shows an input from 0 to 100% before weighting. 3 of 3 listed models have open weights. Most-starred official repository: OpenHelix-Team/VLA-Adapter (2.3k stars, 19 Oct 2026).
The score (0–100) measures public, checkable evidence about a company's robot models. It does not measure private capability. A company that publishes no benchmark results and no weights can still have strong models. Every input is computed from the data on each company page, and every number there links to its source.
| Input | Weight | How it is measured |
|---|---|---|
| Published benchmarks | 35% | For each scored benchmark, the company's best result divided by the best result any company has published, averaged over the benchmarks it reports. No published results scores 0. Results come only from the model's own paper, model card or the official benchmark leaderboard. Scored: LIBERO (4-suite average success, %); SimplerEnv Google Robot, visual matching (%); SimplerEnv Google Robot, variant aggregation (%); SimplerEnv WidowX / Bridge (%); CALVIN ABC→D (average length, 0–5); RoboArena (pairwise rating). |
| Open weights | 20% | Share of the company's listed models whose weights the public can download. |
| Adoption | 20% | GitHub stars on the company's most-starred official model repository, on a log scale relative to the most-starred repository in the list. Hugging Face downloads are shown on company pages but not scored, because gated repositories report zero. |
| Released models | 15% | Number of publicly announced, named models, capped at 5. |
| Disclosed funding | 10% | The sum of the funding rounds listed on the company's profile that were disclosed in USD, on a log scale from $10M to the largest sum in the list. Rounds reported only in other currencies are listed but not converted or counted. This input applies only to private startups and robot makers. For big-tech divisions, research labs and listed companies it is n/a, and its weight is shared across the other inputs in proportion. |
Sources. arXiv papers and official model cards for benchmarks and parameters; official leaderboards (RoboArena); the GitHub and Hugging Face APIs for stars and downloads; company announcements and established press for funding. Missing values are shown as n/a, never estimated.
Updated 19 Oct 2026. Benchmark protocols differ between papers, for example in fine-tuning and number of trials, so small gaps between close scores are not meaningful.
Open weightsVLA
Training method and checkpoints that align VLA intermediate visual embeddings with features from a pretrained 3D foundation model, without depth input at inference.
Open weightsVLA
Tiny-scale VLA that bridges a Qwen2.5-0.5B vision-language backbone to a 97M-parameter policy through bridge attention over raw and action-query features.
| Benchmark | Result | Source |
|---|---|---|
| LIBEROAverage success rate across LIBERO-Spatial/Object/Goal/Long (%) | 97.3% | arxiv.orgTable 5, VLA-Adapter (0.5B). As of 11 Sep 2025 |
| CALVIN ABC→DAverage successful length, CALVIN ABC->D (0-5) | 4.42 | arxiv.orgTable 6, VLA-Adapter; VLA-Adapter-Pro 4.50. As of 11 Sep 2025 |
Open weightsVLA
VLA trained with an auxiliary diffusion objective that reconstructs the gaze region of the target object to improve visual grounding.
| Benchmark | Result | Source |
|---|---|---|
| CALVIN ABC→DAverage successful length, CALVIN ABC->D (0-5) | 3.95 | arxiv.orgTable 3, ABC->D. As of 14 Aug 2025 |
