MID

DexVLA

Midea Group, China · February 2025

Midea Group with East China Normal University and partners

Compact VLAClosed weightsAlso written DexVLA with ScaleDP diffusion expert
Parameters
not published
The paper names a Qwen2-VL-2B backbone and states that the diffusion expert is scaled to one billion parameters, but it never...
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
CC BY-NC 4.0.

What it is

DexVLA keeps a small Qwen2-VL-2B vision language model frozen in role and puts the capacity into the action side, a ScaleDP diffusion expert scaled to about one billion parameters that is pretrained on cross embodiment data before being attached. The three stage embodiment curriculum lets that expert be trained once and then aligned to a specific robot, which is what makes the model cheap to move between arms. It reports 97.3 percent average on LIBERO and completes long horizon tasks such as laundry folding from direct language prompting, and at about 60 Hz on a single A6000 it is by a wide margin the fastest model in this comparison. Two things constrain adoption: the licence is CC BY-NC 4.0, which rules out commercial deployment, and no full DexVLA checkpoint was ever published, only the pretrained diffusion experts, so the VLA has to be trained from the released code. The published LIBERO average also omits the Long suite, which is where most VLAs lose the most.

Architecture

Backbone
Qwen2-VL-2B-Instruct, used off the shelf with no post-training of the VLM itself
Action head
ScaleDP plug-in diffusion expert, released in two sizes: ScaleDP-H at about 1B parameters and ScaleDP-L at about 410M. The expert is pretrained separately on cross embodiment data before being attached to the VLM.
Control rate
60 Hz
Parameters
The paper names a Qwen2-VL-2B backbone and states that the diffusion expert is scaled to one billion parameters, but it never gives a combined total. No official full DexVLA checkpoint is published from which a count could be read, only the two pretrained ScaleDP experts.
Pretraining data
About 100 hours of collected data spanning 91 distinct tasks across four embodiments, with AgileX contributing 42.7 percent and Franka 34.7 percent of the mix.
Embodiments
Franka Emika single arm, AgileX bimanual, Dexterous hand setup, LIBERO (simulation)

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.

Published figures. These come from the model authors, not from this site.
Fine tuning memory, reported38 GB
Control rate60 Hz

Fine-tuning was tested on a single A6000 with 46 GB. Full training needs about 38 GB, LoRA about 24 GB and DeepSpeed offload about 20 to 29 GB, all at batch size 2. No inference VRAM number is published.

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

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO Object
    99.1%
    success rateSame table: pi0 98.8, OpenVLA 88.4, Diffusion Policy 92.5.source
  • LIBERO average over Spatial, Object and Goal only
    97.3%
    success rateThe Long suite is not part of this average, so this number is not comparable with four suite LIBERO averages reported by other models. Same table: pi0 97.1, pi0-FAST 93.9, OpenVLA 84.1.source
  • LIBERO Spatial
    97.2%
    success rateObtained with the stage 1 pretrained diffusion expert weights. Same table: pi0 96.8, OpenVLA 84.7, Diffusion Policy 78.3.source
  • LIBERO Goal
    95.6%
    success rateSlightly below pi0 at 95.8 in the same table.source

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on Franka single arm, AgileX bimanual and a dexterous hand setup. The paper reports long horizon results in which DexVLA scores 0.4 on laundry folding against 0.2 for pi0 and 0.70 on the hard table bussing task against 0.62 for pi0, and it demonstrates direct language prompted shirt folding. Pretraining used about 100 hours of data across 91 tasks and four embodiments. Inference runs at about 60 Hz on a single A6000.

Fine tuning it yourself

Data is H5py in the ACT layout with left, right and wrist camera views plus joint positions and actions. Training is three staged: pretrain the diffusion expert alone on cross embodiment data, align the VLA to the target embodiment, then post-train on the target task. The README reports single A6000 runs at batch size 2 needing about 38 GB for full training, about 24 GB with LoRA, about 20 to 29 GB with DeepSpeed offload and about 33 GB with the smaller ScaleDP. Training only the diffusion expert runs at 0.89 epochs per hour versus 0.32 for the entire VLA.

Where it helps, where it does not

Strengths

  • About 60 Hz on a single A6000, an order of magnitude faster than the diffusion VLAs that report around 5 Hz
  • Real published training VRAM figures, including about 24 GB with LoRA, so fine-tuning fits on one 24 GB card
  • The diffusion expert is separable and pretrained once, then aligned per embodiment, which cuts the cost of adding a new robot
  • Demonstrated on single arm, bimanual and dexterous hand hardware, and on long horizon laundry folding from direct language prompting

Limits

  • CC BY-NC 4.0 licence. Commercial use is explicitly forbidden, which disqualifies the model for most product work.
  • No official full DexVLA checkpoint exists. Only the pretrained ScaleDP-H and ScaleDP-L diffusion experts are released, so the VLA must be trained from the code.
  • The LIBERO average of 97.3 covers only the Spatial, Object and Goal suites. The Long suite, where VLAs typically lose the most, is not evaluated.
  • No inference VRAM figure, no millisecond latency and no documented action chunk size, so the 60 Hz claim cannot be decomposed.
  • Training data is H5py in the ACT layout, not RLDS and not LeRobot, so existing Open X-Embodiment pipelines need conversion.
  • Pretraining is only about 100 hours across 91 tasks, small compared with the million trajectory corpora of foundation VLAs.
  • Repository has been dormant since April 2025.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

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.