A policy plugin for the Solarian framework that trains and serves a
UMI DiffusionUnetTimmPolicy.
Unlike a skill plugin (a behavior), a policy plugin is keyed to a model family/library and ships both halves of a policy:
DiffusionTrainer— asolarian.core.ai.system1.PolicyTrainerthat trains the policy by reusing the UMIdiffusion_policyworkspace (theacceleratemulti-GPU loop). Discovered bysolarian trainthrough thesolarian.trainersentry point (trainer: diffusion).DiffusionPolicyServer— asolarian.core.ai.system1.PolicyServerthat serves a trained checkpoint over HTTP.
The UMI model + training code (the diffusion_policy package — timm workspace,
UmiDataset, model code, and the bundled Hydra configs — plus the small
umi.common.pose_util helper) is vendored directly into this plugin, so
there is no separate UMI install to manage. The HTTP-serving stack,
hydra-core, omegaconf and dill ship with solarian core; only the ML
runtimes (PyTorch, diffusers, timm, …) are optional extras.
# Trainer + server discovery only (no torch):
pip install solarian-diffusion
# Training (accelerate + the diffusion model stack):
pip install "solarian-diffusion[train]"
# Serving a checkpoint:
pip install "solarian-diffusion[serve]"Add a training file, conf/training/<name>.yaml, and set trainer: diffusion:
trainer: diffusion
job_name: cup_diffusion
train_dataset: /data/datasets/cup_in_the_wild.zarr.zip # or local@cup_in_the_wild.zarr.zip
batch_size: 64
num_workers: 4
num_epochs: 120
config_name: train_diffusion_unet_timm_umi_workspace
overrides: [] # raw Hydra overrides, applied last
logger:
type: wandb
project: umisolarian train --conf conf/training/<name>The trainer composes the named UMI workspace config from the vendored
diffusion_policy package, applies the overrides above, and runs
workspace.run() — the same loop as UMI's train.py.
Install solarian-modal and run solarian modal-train --conf conf/training/<name> with
a modal: block (gpu_type, gpus, volume, ...) in the config. It builds
the image from this plugin's modal_spec(), uploads the local dataset the
config names, and runs training remotely with accelerate launch.
solarian-diffusion-policy --input <checkpoint.ckpt> --port 8001Point any skill (e.g. solarian-skill-cup-arrange) at it via
policy_server_url.
Apache License 2.0.
Copyright 2026 PrimaryKey Inc. The vendored diffusion_policy and umi
packages are derived from
Universal Manipulation Interface
and Diffusion Policy (MIT);
see NOTICE.
Solarian and the Solarian logo are trademarks of PrimaryKey Inc. See the trademark policy.