VLA Model Statistics (2026): 50+ Data Points on Model Releases, Training Data, and Benchmark Performance

At least 15 major vision-language-action (VLA) models shipped between December 2022 (Google's RT-1) and mid-2026 (NVIDIA's GR00T N1.7), and ICLR submissions tagged "vision-language-action" jumped 18x in one year — from 9 in 2025 to 164 in 2026. Parameter counts span three orders of magnitude (35M to 55B), yet the 7B OpenVLA beat the 55B RT-2-X by 16.5 points absolute, and disclosed training corpora now run from Figure Helix's ~500 hours of teleoperation to the ~20,000 hours behind Ant Group's LingBot-VLA. Control frequency climbed from 3 Hz to 200 Hz over the same window. We aggregated 59 statistics from arXiv papers and official research published by Google DeepMind, Physical Intelligence, Figure, NVIDIA, Hugging Face, AgiBot, and Ant Group.

01The VLA Release Timeline: 15 Major Models in 43 Months

The release cadence tells the story better than any single model: one major VLA shipped in 2022, seven shipped in 2025. Research volume is compounding even faster — submissions to ICLR with the "vision-language-action" keyword went from 1 in 2024 to 9 in 2025 to 164 in 2026.

By mid-2026 the frontier had moved from lab demos to production. Figure's Helix-powered fleet publicly sorted 22,000 packages over a 17-hour autonomous shift in May 2026, per coverage of the company's livestream.

n/d = not disclosed. "Major release" = a named model with a primary announcement from its developer; incremental checkpoints (GR00T N1.5/N1.6, Helix updates) are folded into their model family. RT-1 predates the "VLA" term, which RT-2 introduced, but is included as the direct architectural ancestor.

02Parameter Counts: From 35M to 55B and Back Down

Scale peaked early. RT-2's 55B PaLI-X variant remains the largest VLA ever disclosed, yet every flagship since has shipped smaller: pi-0 at 3B, GR00T N1 at 2.2B, Helix's reasoning system at 7B. OpenVLA proved the point — 7x fewer parameters than RT-2-X and 16.5 points better. The binding constraint moved from model capacity to data and latency.

Note: RT-1's 35M and Octo-Base's 93M parameter counts are widely cited figures that we could not re-verify against the paper PDFs; treat as reported. GR00T N1's Hugging Face card rounds to "2B"; the paper states 2.2B total.

03How VLA Models Are Trained: The Data Mixtures

Every VLA recipe is a mixture: internet-scale vision-language pretraining, then robot demonstrations, increasingly cut with human egocentric video and synthetic rollouts. The disclosed real-robot corpora span 40x — Helix's ~500 hours to LingBot-VLA's ~20,000 hours.

GR00T N1's mixture shows where recipes are headed: only 88 of its training hours were real in-house teleoperation, supplemented by 827 hours of neural-generated video and 6,500 simulation-hours produced in 11 hours of compute. Real demonstration data remains the scarce, load-bearing ingredient — which is why annotated real-world robot data produced at industrial scale has become a supply chain of its own.

Note: hours are not directly comparable across rows — pi-0.5's ~400 h counts mobile-manipulation data only, GR00T N1.7's 20,000 h is human video rather than robot teleoperation, and SmolVLA trained on fewer than 30k community-contributed episodes.

04Open Datasets Powering VLAs

Four corpora anchor nearly every open VLA: Open X-Embodiment for breadth (22 embodiments), DROID for scene diversity (564 scenes), BridgeData V2 for single-platform depth, and AgiBot World for humanoid-scale volume. Policies pretrained on AgiBot World beat Open X-Embodiment-trained baselines by 30% on average — evidence that dataset coherence and quality, not just size, moves benchmarks. For the full picture of who is producing this data and at what volume, see our robotics training data statistics.

Note: BridgeData V2 (2023) and Open X-Embodiment (2023) figures are the original published counts; DROID's are from its March 2024 paper — the most recent available for each dataset.

05Benchmark Success Rates: What Actually Improved

Three generations of headline numbers: RT-1 hit 97% on instructions it had seen but only 32%-level generalization; RT-2 doubled unseen-scenario success to 62% by inheriting web knowledge; OpenVLA-OFT pushed the LIBERO simulation benchmark to 97.1%. Simulation benchmarks are saturating — above 95% on LIBERO is now table stakes, per the ICLR 2026 VLA analysis.

Real-robot benchmarks are a different story. On Ant Group's 100-task GM-100 real-world benchmark, the best reported model succeeds about 17% of the time (widely reported in launch coverage, though we could not independently re-verify the table) — a reminder that open-world manipulation is far from solved.

Note: the GM-100 figures are as reported in launch coverage of the LingBot-VLA paper; the original results table was not independently re-verified. Benchmark numbers across rows use different task sets and are not directly comparable.

06Inference Speed and Control Frequency

Control frequency is where VLAs became deployable. RT-1 commanded actions at 3 Hz; pi-0's flow-matching head outputs action chunks at up to 50 Hz; Helix's 80M-parameter S1 closes the loop at 200 Hz on embedded low-power GPUs — a ~67x jump in 26 months.

Fine-tuning recipes matter as much as architecture: OpenVLA-OFT multiplied action-generation throughput 26x without touching the base model. Hitting these loop rates in deployment also depends on the sensor stack — camera frame rates, synchronization, and latency budgets — which is where capture hardware matched to control-loop requirements enters the spec sheet.

Note: OpenVLA's ~6 Hz figure is the widely reported single-GPU throughput; models with action chunking (pi-0, OpenVLA-OFT) emit multi-step chunks per inference, so effective control rate exceeds raw model calls per second.

07The Demonstration-Data Bottleneck

Every result above routes through the same constraint: real robot demonstrations are the scarcest input in the stack. The clearest published dose-response curve comes from Figure — six times the demonstration hours (10 → 60) cut per-package handling time 37% and pushed barcode success from 88.2% to 94.4%.

The scaling-laws evidence says diversity beats raw volume. A 40,000-demonstration study found generalization follows a power law in the number of distinct environments and objects — not demonstration count — and pi-0.5 needed only ~100 distinct environments to approach a policy trained directly on the test home. That reframes the procurement question from "how many hours" to "how many scenes," the logic behind spec-to-pilot-to-scale data production across many real environments. For collection methods, throughput, and cost per hour, see our teleoperation and demonstration data collection statistics.

Note: the scaling-laws study also found that beyond a per-environment threshold, additional demonstrations in the same environment add little — the marginal value is in new scenes and objects.

Summary: VLA Models by the Numbers

Methodology and Sources

This article aggregates 59 statistics from 24 primary sources: peer-reviewed and preprint papers on arXiv, official research blogs and model cards from Google DeepMind, Physical Intelligence, Figure AI, NVIDIA, Hugging Face, AgiBot, and Ant Group, and one OpenReview-based analysis of ICLR submission data. We cite original research directly rather than secondary roundups. Stats we could not trace to a fetchable primary document (RT-1's 35M and Octo's 93M parameter counts, AgiBot World Beta's 2,976.4 hours, the GM-100 benchmark table, and Figure's 22,000-package shift) are flagged inline as reported figures. Historical entries (2022–2023 models and datasets) are dated in every table.

• Google Research — RT-1: Robotics Transformer (2022)
• RT-1 paper — arXiv 2212.06817 (2022)
• Google DeepMind — RT-2 announcement (2023)
• Open X-Embodiment — arXiv 2310.08864 (2023)
• Open X-Embodiment / RT-X project page (2023)
• UC Berkeley — BridgeData V2 (2023)
• Octo: An Open-Source Generalist Robot Policy — arXiv 2405.12213 (2024)
• OpenVLA — arXiv 2406.09246 (2024)
• DROID — arXiv 2403.12945 (2024)
• Physical Intelligence — pi-0 (2024)
• Physical Intelligence — openpi repository (2025)
• Physical Intelligence — pi-0.5 (2025)
• Data Scaling Laws in Imitation Learning — arXiv 2410.18647 (2024)
• Figure AI — Helix announcement (2025)
• Figure AI — Scaling Helix in logistics (2025)
• NVIDIA — GR00T N1 paper, arXiv 2503.14734 (2025)
• NVIDIA — Isaac GR00T repository, N1.7 (2026)
• Google DeepMind — Gemini Robotics paper, arXiv 2503.20020 (2025)
• Google DeepMind — Gemini Robotics On-Device (2025)
• Google DeepMind — Gemini Robotics 1.5 (2025)
• Hugging Face — SmolVLA (2025)
• AgiBot — AgiBot World Colosseo, arXiv 2503.06669 (2025)
• AgiBot World — Hugging Face dataset organization (2025)
• Ant Group — LingBot-VLA, arXiv 2601.18692 (2026)
• OpenVLA-OFT — arXiv 2502.19645 (2025)
• M. Reuss — State of VLA research at ICLR 2026
• eWeek — Figure Helix 22,000-package shift coverage (2026)

Last updated July 2026. We update this page quarterly as new data becomes available.