VLA Gallery / All models / Industry models

RynnVLA-001

Alibaba DAMO Academy · 2025-08

VLAIndustry models7Blatent actions

Three-stage training: pretrains video generation on 12M egocentric human manipulation videos, then trajectory-aware video modeling and VLA fine-tuning.

Paper

RynnVLA-001: Using Human Demonstrations to Improve Robot ManipulationarXiv 2509.15212

Overview

Category

Industry models

Type

VLA

Released

2025-08

Organization

Alibaba DAMO Academy

Open weights

Yes

Parameters

7B (public model names RynnVLA-001-7B-Base · 7B-Trajectory; paper total not disclosed)

Architecture

Backbone LLM

Chameleon 7B (video extension)

Vision encoder

Chameleon image tokenizer

Action head

ActionVAE latent actions (latent actions)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Extends the Chameleon 7B image generation model to video (README initial weights 'Chemeleon-7B'); ActionVAE compresses action sequences into latent embeddings

Training

Pretraining accelerators

Not disclosed (only ActionVAE uses 8 GPUs, model not specified)

Pretraining duration

Not disclosed

Pretraining steps & batch

ActionVAE 300,000 iterations, batch 16×8(GPUs)

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed

Data type

Human video + robot video + motion capture (EgoDex upper-body joint trajectories) + real-robot demonstrations (SO-100)

Data scale

12M egocentric human manipulation videos + 244K robot manipulation videos; fine-tuning 248·249·301 demos

Main datasets

In-house ego-centric video, EgoDex (Apple Vision Pro), in-house SO-100 demonstrations

Tasks & evaluation

SO-100 3 tasks (block pick-and-place, strawberry, pen insertion) average 90.6% vs π0 70.4% vs GR00T N1.5 55.6% (fine-tuned on the same data)

Robot

LeRobot SO-100 single arm (3 units)

Inference

Inference device & speed

Not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

14 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Video generation pretraining + imitation learning

Base control

Fixed · No base control (SO-100 single arm)

Notes

Key features

Three-stage training: pretrains video generation on 12M egocentric human manipulation videos, then trajectory-aware video modeling and VLA fine-tuning. Averaged 90.6% on 3 SO-100 tasks, ahead of π0 (70.4%) and GR00T N1.5 (55.6%).

Notes

Weights·code released 2025-08-08 (Apache 2.0), technical report 2025-09-18. Parameters (B) based on public model name.

Verification

Confirmed