RynnVLA-001
Alibaba, China · August 2025
Alibaba DAMO Academy
- Parameters
- 7 B
- Both released checkpoints are 7B: RynnVLA-001-7B-Base (initialized from Chameleon-7B, after ego-centric video generative...
- GPU memory
- 13 GB
- weights at bf16, computed. Fits a 24 GB card
- Inference latency
- not published
- per action step
- Weights
- Apache-2.0
- Apache-2.0 (the repository additionally describes the release as a research preview intended for non-commercial use only)
What it is
RynnVLA-001 is Alibaba DAMO Academy's VLA built on Chameleon-7B, and its thesis is that human demonstrations in ego-centric video are the right pretraining substrate for robot manipulation. Stage one trains an image-to-video model on 12 million ego-centric manipulation videos, stage two adds joint prediction of future keypoint trajectories to bridge frame prediction and action prediction, and an ActionVAE compresses action chunks into compact latent embeddings so the policy output space stays small. On a LeRobot SO-100 arm it reports 90.6 percent average success against 70.4 percent for pi-0 and 55.6 percent for GR00T N1.5. One number deserves attention: first-attempt success (SR@1) is 56.7 percent against pi-0's 56.3 percent, essentially a tie, so much of the headline gap comes from recovery behaviour rather than getting it right the first time.
Architecture
- Backbone
- Chameleon-7B, an autoregressive unified image and text transformer, extended to jointly predict next frames and next actions
- Action head
- ActionVAE, a variational autoencoder that compresses action chunks into compact continuous latent embeddings, with two domain-specific VAEs trained separately, one for human trajectories and one for robot actions
- Parameters
- Both released checkpoints are 7B: RynnVLA-001-7B-Base (initialized from Chameleon-7B, after ego-centric video generative pretraining) and RynnVLA-001-7B-Trajectory (initialized from the Base checkpoint, after trajectory-aware video modeling). The ActionVAE is a separate lightweight module whose parameter count is not published.
- Pretraining data
- Two-stage pretraining. Stage 1, Ego-Centric Video Generative Pretraining, trains an image-to-video model on 12 million ego-centric manipulation videos to predict future frames from an initial frame and a language instruction. Stage 2, Human-Centric Trajectory-Aware Modeling, adds joint prediction of future keypoint trajectories to bridge frame prediction and action prediction.
- Embodiments
- LeRobot SO-100 single-arm robot with a front camera and a wrist camera
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 26 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 13 GB | The usual way these checkpoints are served |
| int8 | 6.5 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 3.3 GB | Aggressive quantisation, verify success rate before trusting it |
Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.
No VRAM figure is published. The only hardware datapoint in the repository is that the ActionVAE was trained for 300000 iterations with a batch size of 16 across 8 GPUs, which does not translate into a requirement. A 7B model in BF16 needs roughly 14GB for weights alone, but since no source states an inference figure none is recorded.
The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
No results in this category are published for this model.
Real world
Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.
- Real robot (LeRobot SO-100) instruction following with distractors91.7%success rateAuthors' own evaluation. pi-0 scores 60.0 percent and GR00T N1.5 scores 56.7 percent. This is the setting with the largest margin.source
- Real robot (LeRobot SO-100) average over three manipulation tasks90.6%success rateAuthors' own evaluation. pi-0 scores 70.4 percent and GR00T N1.5 scores 55.6 percent on the same hardware and tasks.source
- Real robot (LeRobot SO-100) grab pen and put it into holder (hardest single task)90%success rateAuthors' own evaluation. pi-0 scores 64.4 percent and GR00T N1.5 scores 48.3 percent.source
- Real robot (LeRobot SO-100) multi-target manipulation86.7%success rateAuthors' own evaluation. pi-0 scores 71.1 percent, GR00T N1.5 scores 46.7 percent.source
- Real robot (LeRobot SO-100) average over three tasks, first-attempt success56.7%SR@1 (success rate at first attempt)Authors' own evaluation. pi-0 scores 56.3 percent, effectively a tie. This row qualifies the 90.6 percent headline substantially.source
Fine tuned tasks
No results in this category are published for this model.
On real hardware
Evaluated on a LeRobot SO-100 arm with front and wrist cameras. Average success rate 90.6 percent against 70.4 percent for pi-0 and 55.6 percent for GR00T N1.5 across three tasks. Broken down by setting: 93.3 percent single-target, 86.7 percent multi-target and 91.7 percent instruction following with distractors.
Fine tuning it yourself
Training data is supplied as LeRobot-format episodes converted to HDF5. Two checkpoints are released so you can start either from the video-pretrained Base or from the trajectory-aware checkpoint. The licence situation needs care: the repository states Apache-2.0 but also describes the release as a research preview for non-commercial use only, so legal review is advisable before commercial deployment.
Where it helps, where it does not
Strengths
- Runs on the LeRobot SO-100, a low-cost open-hardware arm, so the real-robot results are reproducible on a hobby budget rather than on a proprietary platform, which is unusual in this comparison
- Clear margin over pi-0 and GR00T N1.5 on the same hardware and the same tasks (90.6 versus 70.4 and 55.6 percent average)
- Largest single margin is on instruction following with distractors (91.7 versus 60.0 percent for pi-0), which is the setting most VLAs fail
- Open weights, two checkpoints covering both pretraining stages
- Accepted at ICRA 2026
Limits
- No published parameter-efficient training path, no VRAM figure and no latency, on a 7B model, so deployment sizing has to be measured
- Licence is ambiguous: Apache-2.0 in the repository but simultaneously described as a research preview for non-commercial use only
- Evaluated only on a single-arm SO-100 with three tasks, so no bimanual, mobile or long-horizon evidence exists
- No simulation benchmark results are reported in the paper, so there is no LIBERO or CALVIN number to compare against other models
- SR@1 (first-attempt success) is only 56.7 percent against 56.3 percent for pi-0, essentially a tie, so the headline margin comes largely from retry behaviour rather than first-shot accuracy
- Superseded by RynnVLA-002 in November 2025
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://arxiv.org/abs/2509.15212
- https://arxiv.org/html/2509.15212v1
- https://github.com/alibaba-damo-academy/RynnVLA-001
- https://huggingface.co/blog/Alibaba-DAMO-Academy/rynnvla-001
- https://huggingface.co/Alibaba-DAMO-Academy
Entry last checked 2026-08-11.
Try a policy on a real arm
A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.