ChatVLA-2
Midea Group, China · May 2025
Midea Group and East China Normal University
- Parameters
- not published
- The paper does not state a total parameter count.
- GPU memory
- not published
- weights at bf16, computed
- Inference latency
- not published
- per action step
- Weights
- closed
- Not stated.
What it is
ChatVLA-2 is the authors' own successor to ChatVLA and shifts the goal from retaining question answering ability to retaining open world reasoning that actually drives the arm. It replaces the dual expert design with a dynamic mixture of eight experts, two active per token, and attaches a pretrained 1B ScaleDP diffusion expert through a small projection module. The evaluation is deliberately adversarial: a math matching game with unseen numbers, symbols and backgrounds, and a toy placement task with unseen objects, where it reaches 43 of 52 and 127 of 156 successes against 4 of 52 and 22 of 156 for ChatVLA and 8 of 52 and 25 of 156 for pi0. The catch for anyone planning to deploy it is that nothing has been released. No weights, no code repository and no license were found as of August 2026, and no parameter count, VRAM, latency or control frequency figure is published, so the model exists as a NeurIPS 2025 result rather than as something you can run.
Architecture
- Backbone
- Qwen2-VL based vision language model with a dynamic mixture of experts using eight experts and two active per token
- Action head
- A pretrained 1B parameter ScaleDP diffusion expert, connected to the VLM through a projection module of two linear layers and a LayerNorm
- Parameters
- The paper does not state a total parameter count. It names a 1B ScaleDP action expert and an eight expert mixture of experts over a Qwen2-VL backbone, and it uses 3B and 7B dense models only as comparison baselines. No checkpoint exists from which a count could be read.
- Pretraining data
- Co-training mixture of image text data (COCO 32K, TextVQA 20K, GQA 54K) with robot demonstrations (600 trajectories for math matching, 300 for toy placement) at an image text to robot ratio of 1 to 3. Total training cost is reported as 340 GPU hours.
- Embodiments
- ARX R5 bimanual robot (14 DoF), Franka Emika 7-DoF arm with a Robotiq gripper
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.
No inference GPU or VRAM figure is published. The only compute statement is 340 GPU hours of total training, without naming the GPU type.
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.
- Toy placement (authors' open-world real robot task) 156 trials, unseen objects and targets127 of 156 trials
- Math matching game (authors' open-world real robot task) 52 trials, unseen numbers, answers, backgrounds and symbols43 of 52 trialssuccessful trialsAuthors' own task, not a public benchmark. Same table: DexVLA 10 of 52, pi0 8 of 52, ChatVLA 4 of 52.source
- Math matching game (authors' open-world real robot task) 52 trials, OCR sub-score3.58 of 4 points
- Toy placement (authors' open-world real robot task) 156 trials, object recognition sub-score0.94 normalised score
Fine tuned tasks
No results in this category are published for this model.
On real hardware
All reported results are real world. Two setups are used: an ARX R5 bimanual robot with 14 DoF and a Franka Emika 7-DoF arm with a Robotiq gripper, both with a RealSense L515 camera. Training used 600 trajectories for the math matching task and 300 for toy placement.
Fine tuning it yourself
Co-training mixes image text data (COCO, TextVQA, GQA) with robot demonstrations at a 1 to 3 ratio. Total training cost is reported as 340 GPU hours. No public training code was found, so this pipeline is not reproducible from released artifacts.
Where it helps, where it does not
Strengths
- Large measured margin over ChatVLA, DexVLA and pi0 on deliberately open-world tasks with unseen numbers, symbols and objects
- Reasoning is evaluated as reasoning that reaches the actuator, not only as benchmark question answering
- Reports its own training cost of 340 GPU hours, which is unusually transparent
- Evaluated on two distinct embodiments including a 14 DoF bimanual ARX R5
Limits
- No weights, no code repository and no license could be found as of August 2026. The model cannot be run or reproduced.
- No total parameter count, no VRAM, no latency, no control frequency and no action chunk size are published.
- Evaluation rests on two custom real world tasks on the authors' own hardware. There is no LIBERO, SIMPLER or CALVIN number for cross comparison.
- Robot training data is small, 600 and 300 trajectories per task, so the reported generalisation applies to variations of two task families rather than to a broad skill set.
- Only 340 GPU hours of training, which limits how much of the claimed capability can come from scale rather than from the specific task setup.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://arxiv.org/abs/2505.21906
- https://chatvla-2.github.io/
- https://github.com/tutujingyugang1/ChatVLA_public
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.