CU

Diffusion Policy (Transformer-based, DiffusionPolicy-T)

Columbia University, United States · March 2023

Columbia University, with Toyota Research Institute and MIT

Action policyClosed weightsAlso written DP-T, DiffusionPolicy-T, Time-series diffusion transformer
Parameters
31 M
About 31 M for the image based Robomimic configuration: 9 M for the diffusion transformer plus 22 M for the vision encoder,...
GPU memory
0.1 GB
weights at bf16, computed. Fits an 8 GB card
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

DiffusionPolicy-T replaces the 1D temporal convolutional denoiser of the CNN variant with a time-series diffusion transformer that cross-attends to the observation embedding. On the paper's state-based Robomimic splits it reaches the highest scores of any configuration on Lift, Can and Square, and it is the stronger option on the multi-stage BlockPush and Franka Kitchen tasks. On the harder long-horizon tasks it falls behind the CNN variant, losing 4 points on state-based ToolHang. Since the paper reports no parameter count and no latency figure for this variant, hardware sizing for it has to come from the code, not from the publication.

Architecture

Backbone
No VLM and no language input. A vision encoder produces an observation embedding that conditions a time-series diffusion transformer, which denoises the action sequence with cross-attention to the observation embedding.
Action head
Conditional denoising diffusion (DDPM) over an action chunk, transformer denoiser instead of the U-Net used by the CNN variant, DDIM at inference.
Control rate
10 Hz
Parameters
About 31 M for the image based Robomimic configuration: 9 M for the diffusion transformer plus 22 M for the vision encoder, roughly one ninth of the CNN variant. Transport uses a larger encoder (9 M plus 45 M).
Pretraining data
None. Trained per task from demonstrations, same evaluation protocol as the CNN variant.
Embodiments
Robomimic simulation tasks Lift, Can, Square and ToolHang (single arm) and Transport (two arms), Push-T, simulated, Franka Kitchen and BlockPush 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.

Computed from 31 M parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp320.1 GBTraining master weights, rarely used for inference
bf16 or fp160.1 GBThe usual way these checkpoints are served
int80.0 GBQuantised, expect some loss of precision on fine motions
int40.0 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 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.

Published figures. These come from the model authors, not from this site.
Control rate10 Hz

The paper does not state a minimum GPU or a VRAM figure, and the 0.1 s on RTX 3080 latency figure is reported for the real-world deployment without attributing it specifically to the transformer variant. Do not transfer the CNN variant's latency figure to this row.

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.

  • Robomimic Square, proficient human demonstrations, image observations
    90%
    success rateAverage of the last 10 checkpoints, the second number in each cell of the paper table.source
  • Robomimic Square, proficient human demonstrations, state observations
    89%
    success rateAverage of the last 10 checkpoints, the second number in each cell of the paper table.source
  • Robomimic ToolHang, proficient human demonstrations, state observations
    87%
    success rateAverage of the last 10 checkpoints. Far ahead of the CNN variant here (30), and behind it on the image split.source
  • Robomimic Transport, proficient human demonstrations, state observations
    84%
    success rateAverage of the last 10 checkpoints, the second number in each cell of the paper table.source
  • Robomimic Transport, proficient human demonstrations, image observations
    81%
    success rateAverage of the last 10 checkpoints, the second number in each cell of the paper table.source
  • Push-T state observations
    79%
    success rateAverage of the last 10 checkpoints, the second number in each cell of the paper table.source
  • Push-T image observations
    66%
    success rateAverage of the last 10 checkpoints, the second number in each cell of the paper table.source
  • Robomimic ToolHang, proficient human demonstrations, image observations
    47%
    success rateAverage of the last 10 checkpoints. The CNN variant reaches 73 on this split.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

The paper's real-world deployment numbers (0.1 s inference on an RTX 3080, 10 Hz policy rate interpolated to 125 Hz) are reported for the real-world experiment setup without being attributed to the transformer variant specifically, so no real-world figure is claimed for this row.

Fine tuning it yourself

Same pipeline as the CNN variant: per-task training from a zarr demonstration dataset, no pretrained trunk. Selected in the repository by choosing the transformer denoiser config instead of the U-Net config.

Where it helps, where it does not

Strengths

  • Reaches 100 percent average-of-last-10-checkpoints success on state-based Robomimic Lift, Can and Square, the strongest numbers in the paper's state-policy table.
  • Handles high-dimensional action spaces and multi-stage tasks well, which is where the paper's BlockPush and Franka Kitchen results separate it from earlier baselines.
  • Same MIT licensed codebase and same training pipeline as the CNN variant, so switching between the two is a config change.

Limits

  • Less stable than the CNN variant on the harder splits. It loses 2 points on state-based Transport and 4 points on state-based ToolHang against DiffusionPolicy-C.
  • The paper does not publish a latency measurement, a parameter count or a VRAM figure for this variant specifically.
  • Same structural constraints as the CNN variant: no pretraining, no language conditioning, one model per task, and an iterative denoising loop at inference.
  • More sensitive to hyperparameters than the CNN variant in the authors' benchmark tables, particularly on multi-human demonstration splits.

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.