CogACT-Small

Microsoft, United States · November 2024

Microsoft Research

Foundation VLAOpen weights, MITAlso written CogACT DiT-S
Parameters
not published
The paper gives 7.6B only for the default DiT-Base configuration and never states a total for the Small variant.
GPU memory
30 GB
reported by the authors
Inference latency
not published
per action step
Weights
MIT
downloadable checkpoint

What it is

CogACT-Small is the smallest of the three released CogACT checkpoints and differs from the Base variant only in the diffusion action module, which shrinks from 89M to 13M parameters while the roughly 7B parameter Prismatic backbone stays the same. Because the backbone dominates, the checkpoint is still a 30.22 GB fp32 file and the memory requirement is essentially identical to Base, so this variant buys almost no deployment savings. The paper's action module ablation shows it costs real accuracy: 58.5 percent average across the three SIMPLER settings against 62.5 percent for Base and 64.8 percent for Large. Its numbers come from the ablation table rather than the main results tables, and no real robot evaluation of this variant is published.

Architecture

Backbone
Prismatic VLM at 7B scale: DINOv2 ViT-L/14 plus SigLIP ViT-So400M/14 vision encoders with a Llama-2 language module
Action head
Diffusion transformer (DiT-Small, 13M parameters) conditioned on a single cognition token, sampled with DDIM and classifier-free guidance
Action chunk
16 steps per forward pass
Parameters
The paper gives 7.6B only for the default DiT-Base configuration and never states a total for the Small variant. The action module itself is 13M parameters (Table 7). The released checkpoint file is 30.22 GB in fp32, which is 0.30 GB smaller than the Base checkpoint, consistent with the 76M parameter difference between DiT-S and DiT-B.
Pretraining data
A subset of Open X-Embodiment, about 0.4M trajectories according to the project page.
Embodiments
Google Robot (SIMPLER simulation), WidowX / Bridge (SIMPLER 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.
Inference memory, reported30 GB
Checkpoint download30.22 GB

No variant specific figure. The family wide statement is about 30 GB in fp32 for inference, with a bf16 path mentioned but not measured.

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.

  • SIMPLER Google Robot, visual matching
    73.3%
    success rateFrom the authors' action module ablation table, not the main results table.source
  • SIMPLER average over Google Robot visual matching, Google Robot variant aggregation and WidowX visual matching
    58.5%
    success rateAuthors' own three setting average. Base 62.5, Large 64.8 in the same table.source
  • SIMPLER Google Robot, variant aggregation
    51.3%
    success rateFrom the authors' action module ablation table. Base reaches 61.3 in the same table.source
  • SIMPLER WidowX / Bridge, visual matching
    51%
    success rateFrom the authors' action module ablation 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

No real robot evaluation of the Small variant is published. The real hardware experiments in the paper (Realman and Franka) use the default DiT-Base configuration.

Fine tuning it yourself

Identical pipeline to CogACT-Base: RLDS or Open X-Embodiment data with 7-DoF end effector delta actions, Python 3.10, PyTorch 2.2 or newer, CUDA 12 or newer, Flash-Attention 2 for training, reference config of one node with 8 A100 GPUs.

Where it helps, where it does not

Strengths

  • Same MIT license and same download path as the other two variants
  • Smallest action module in the family, useful as a control point when studying action head scaling

Limits

  • Almost no practical benefit over CogACT-Base: the 76M parameter saving is invisible next to the roughly 7B backbone and the checkpoint is only 0.30 GB smaller.
  • Measurably worse than Base on SIMPLER: 51.3 versus 61.3 on Google Robot variant aggregation.
  • Only ablation table numbers exist for this variant. No real robot results and no separately measured latency were published.
  • No published total parameter count for this specific variant.
  • Same single camera, single arm restriction as the rest of the family.

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.