Open-weights release, September 2026

Griffin AlphaA robust and extensible robot foundation model.

Tan Yi Xuan (LinkedIn, opens in a new tab), Bryce Goh (LinkedIn, opens in a new tab), Yizhuo Zhang (LinkedIn, opens in a new tab), Ang Yan Kang (LinkedIn, opens in a new tab), Tej Deep Pala (LinkedIn, opens in a new tab), Soujanya Poria (LinkedIn, opens in a new tab), Raphael Yee (LinkedIn, opens in a new tab), Griffin Labs

Griffin Alpha is a robot foundation model built for different robots, long tasks and changing environments. We’re releasing its open core, Griffin Alpha-S, alongside CleanBench, our benchmark for real-world cleaning tasks.

Hugging Face GitHub

One set of released Griffin Alpha-S weights running CleanBench on bimanual Piper arms.

Our belief

What a deployable robot foundation model has to be

We build robots for facilities work. That sets three requirements for a model we can deploy.

Embodiment-agnostic by construction

Robots differ in kinematics, cameras and proprioception. The knowledge that transfers between them, how objects behave, how to approach a grasp, what a failing task looks like, is not specific to any one machine, so the model should not be either.

Stable over long horizons

Real work is minutes of dependent steps. The fifth action only makes sense given the first four, and one early mistake invalidates everything after it. A skill that lasts a few seconds demonstrates a component of the job rather than the job.

Robust to a change of domain

Lighting shifts, clutter accumulates, backgrounds change and objects sit a few centimetres off. A policy that loses most of its performance under those changes cannot be left to run a shift.

Approach

How we built it

Griffin Alpha pairs a Qwen3-VL 4B vision-language backbone with a flow-matching action expert. One set of weights supports multiple robots and both joint-space and end-effector actions.

Two-stage architecture: next-token pre-training, followed by a flow-matching action expert that denoises action chunks over K Euler steps.

Stage one: next-token prediction. We extend Qwen3-VL 4B¹¹ with action and proprioception tokens, then pre-train it to predict subtasks and action chunks from images, instructions and robot state^1,2.

Stage two: continuous actions. A flow-matching expert^3,10 reads the backbone’s per-layer keys and values alongside its own.

A general pre-training mixture

Pre-training combines open-source robot corpora⁸ with our own data: dual-arm mobile manipulators, single-arm platforms, simulation and egocentric video. We weight each corpus for the variety of robots, tasks, rooms and lighting it adds.

Pre-training mixture by embodiment

Bimanual (simulation)42.5%

Bimanual (teleop)33.8%

Egocentric9.9%

Single arm (simulation)8.9%

Single arm (teleop)4.9%

Total100%

Share of pre-training data by embodiment. Bimanual data, simulated and teleoperated, makes up three quarters of the mixture. The remainder is there for the variation it introduces.

Results

Griffin Alpha-S on public benchmarks

Griffin Alpha-S evaluated on LIBERO⁶ and RoboTwin 2.0⁷.

Public simulation

LIBERO

On LIBERO’s long-horizon suite, the hardest of the four, Griffin Alpha-S scores 94.0 to π_0.5’s² 93.0.

On the four-suite average it ties π_0.5 at 96.8, level with the state of the art.

ModelLong-horizon suite success%

Griffin Alpha-S94.0

π_0.593.0

NORA-1.589.6

π₀85.2

NORA63.0

π₀-FAST60.2

Baseline results are author-reported^1,2,4,5; small differences should not be treated as definitive rankings.

Public simulation, domain shift

RoboTwin 2.0

On RoboTwin 2.0⁷ clean2clean, Griffin Alpha-S scores 59.8%: 8th of 11 VLAs and eleven points behind π_0.5⁹.

Under domain shift (clean2random), it scores 46.2%, first among those 11 VLAs and level with π_0.5 at 46.0%. This setting tests changes in the scene, clutter and lighting.

Model (VLAs only)clean2random (hard) success%

Griffin Alpha-S46.2

π_0.546.0

OLA-Geo33.6

Abot-M030.4

Spatial Forcing26.7

X-VLA20.9

Xiaomi Robotics-018.2

π₀16.3

EventVLA15.7

GalaxeaVLA12.7

starVLA3.2

All 11 vision-language-action models on clean2random (hard). The public board carries 20 entries; we show only the VLAs, so the comparison stays within one model class. Excluded: the world action models 4D-WAM, X-WAM, FastWAM and AHA-WAM; the non-VLA baselines DP, DP3, ACT and RDT; and GigaBrain-0.7 and OLA-Sem, which held the top two places on this metric at 67.9% and 67.6%. Leaderboard figures from the RoboTwin 2.0 leaderboard, as of 10 September 2026; Griffin Alpha-S figures from our own run.

RoboTwin 2.0, a second reading

Robustness to domain shift

Alpha-S loses 22.7% of its clean-scene score under domain shift, versus 34.9% for π_0.5. This is the smallest relative drop among the 11 VLAs shown; excluded entries GigaBrain-0.7 and OLA-Sem lose less.

Model (VLAs only)0255075100%Drop, ptsDrop, rel.

Griffin Alpha-S−13.6−22.7%

π_0.5−24.7−34.9%

Abot-M0−27.0−47.0%

OLA-Geo−49.4−59.5%

π₀−30.1−64.8%

Spatial Forcing−50.5−65.4%

X-VLA−47.1−69.3%

Xiaomi Robotics-0−44.7−71.1%

EventVLA−49.9−76.1%

GalaxeaVLA−50.0−79.7%

starVLA−43.3−93.1%

clean2clean (easy)clean2random (hard)Performance lost

Drop from clean2clean to clean2random, in points and as a share of the clean2clean score, sorted by relative drop. All 11 vision-language-action models on the board. The four world action models, the four non-VLA baselines and GigaBrain-0.7 and OLA-Sem are excluded, so every row is one model class; the last two give up less than any row shown. Figures from the RoboTwin 2.0 leaderboard, as of 10 September 2026; Griffin Alpha-S figures from our own run. Drops are single-seed differences and carry wider error than the success rates behind them.

CleanBench

Introducing CleanBench

Three cleaning tasks on bimanual Piper arms, scored on the quality of the finished job.

Bimanual Piper and Mobile ALOHA account for 2% of pre-training samples, all simulated. The physical robots, configurations and cameras used here were absent from pre-training.

Clearing a sink countertop

Rigid objects, on arms the model saw only in simulation and in a different configuration. Pick up each item, put it where it belongs, leave the counter clear.

Folding clothes

Deformable objects. A shirt has no canonical pose and its appearance changes as it is handled, so no two attempts look alike.

Clearing a dining table

Long horizon on a cluttered tabletop, and the one task that needs both arms together: objects are passed from one arm to the other, and the policy has to track what it has already moved.

Score, by task

Task and how it is scoredUnit

Sink countertop cleanup

Attempts passed, of 20

20 attempts. A pass requires every item in its designated place and a clear counter.

Griffin Alpha-S20 / 20

π_0.520 / 20

π₀-FAST15 / 20

Clothes folding

Attempts passed, of 20

20 attempts. A pass requires a neatness score of at least 4 out of 5.

Griffin Alpha-S18 / 20

π_0.514 / 20

Clear a dining table

Points, of 15 · Time to clear

20 attempts, scored out of 15: one point per correctly placed object, plus points for handoffs and emptying dishes before racking. We report mean score and mean time from first motion to final placement, capped at three minutes.

Points, of 15 · higher is better

Griffin Alpha-S12.6 / 15

π_0.512.8 / 15

Time to clear, of 3:00 · lower is better

Griffin Alpha-S2:15

π_0.52:43

Identical initial placements, fine-tuning recipe, data and evaluation across models; π₀-FAST tested on the countertop task only. Single seed. Two 20/20 results cannot be distinguished at this sample size. Time bars use a three-minute scale (135.0 s and 162.7 s).

Countertop precision. Both models complete all attempts. On held-out episodes, Griffin Alpha-S tracks trajectories to within 1.5% of joint range, more than halving error against both baselines. It also trains faster than π_0.5 under the same recipe and hardware.

Table-clearing speed. Scores are similar: 12.6/15 for Griffin Alpha-S and 12.8/15 for π_0.5. Alpha-S finishes in 2:15 versus 2:43—17% less time—with fewer regrasps and less hesitation in the recordings.

We’ll expand CleanBench with more tasks and use recorded failures to guide model development.

What’s next

Extending Alpha into World Action Models

All results here use Griffin Alpha-S: the backbone and action expert. Our World Action Models extend that core for deployment.

World models, sensor encoders, planners and memory can attach to the same backbone. The modules change with the job; the core weights remain fixed.

Griffin Alpha stays fixed while modules attach as extra streams into the same token sequence.

Modules attach to the shared backbone as additional input streams. Colours trace each module through the token sequence and connections.

We make deployable robots

Let Griffin Alpha work for you.

We adapt Griffin Alpha and robot hardware to your facility’s tasks. Get in touch to discuss a deployment.

Help build the next model, benchmark and robot deployment. We’re hiring engineers and researchers to work on real hardware.

We’re hiring

Acknowledgements

Special thanks to Google Cloud Platform team for their support, Annie and James.

References

01Black, K. et al. π0: A vision-language-action flow model for general robot control. Physical Intelligence, 2024. arXiv:2410.24164.
02Physical Intelligence et al. π0.5: A vision-language-action model with open-world generalization. 2025. arXiv:2504.16054.
03Pertsch, K. et al. FAST: Efficient action tokenization for vision-language-action models. 2025. arXiv:2501.09747.
04Hung, C.-Y. et al. NORA: A small open-sourced generalist vision language action model for embodied tasks. 2025. arXiv:2504.19854.
05Hung, C.-Y. et al. NORA-1.5: A vision-language-action model trained using world model- and action-based preference rewards. 2025. arXiv:2511.14659.
06Liu, B. et al. LIBERO: Benchmarking knowledge transfer for lifelong robot learning. NeurIPS 2023. arXiv:2306.03310.
07Chen, T. et al. RoboTwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation. 2025. arXiv:2506.18088.
08Open X-Embodiment Collaboration. Open X-Embodiment: Robotic learning datasets and RT-X models. 2023. arXiv:2310.08864.
09RoboTwin Platform. RoboTwin 2.0 leaderboard. robotwin-platform.github.io/leaderboard (opens in a new tab). Accessed 10 September 2026.
10Lipman, Y. et al. Flow matching for generative modeling. ICLR 2023. arXiv:2210.02747.
11Qwen Team. Qwen3-VL technical report. Alibaba, 2025. arXiv:2511.21631.