HomeSpecsDocsStore
Log inSign up
TwitterDiscordGitHub
Skip to documentation
Sourccey docs

Explore the documentation

Getting startedWelcome to SourcceyChoose your path from first setup to policy development.Getting startedBefore you beginPrepare your robot, workspace, computers, and network.Vulcan StudioDownload & installInstall the desktop application for Windows, macOS, or Linux.Developer guidesDeveloper quickstartSet up the controller and robot computers with Python and LeRobot.Developer guidesTeleoperate SourcceyVerify the network, leader arms, robot, keyboard, and cameras together before collecting data.Developer guidesRecord a datasetCollect a small, consistent set of demonstrations after teleoperation passes.Developer guidesReplay an episodeSend recorded actions back to Sourccey and validate the dataset before training.Developer guidesTrain a policyTrain an ACT policy from the dataset without connecting to the robot.Developer guidesRoll out a policyGive a trained policy supervised control of Sourccey for a short test.Developer guidesTroubleshootingWork from the top of the relevant section and run one diagnostic at a time.Hardware & referenceAssembly overviewNine build stages; detailed illustrated instructions are pending.AI agentsAgent quickstartCopy task context and read documentation as Markdown.
Tab to move between results Esc to close
Developer guides

Train a policy

Train an ACT policy from the dataset without connecting to the robot.

01

Choose a device

Check that this guide matches your installed software. If you are unsure which version you have, the team can help you find it.

  • Use cuda for a supported NVIDIA GPU.
  • Use mps for Apple silicon.
  • Use cpu for a CPU-only run.
  • Lower batch_size if the device runs out of memory.
02

Train an ACT policy

Run training on the controller or another computer that can access the dataset. sourccey-host is not required.

Training computer · terminal
uv run lerobot-train \
  --dataset.repo_id=vulcan-studio/sourccey-demo-1 \
  --policy.type=act \
  --policy.device=cuda \
  --output_dir=outputs/train/act-sourccey-demo-1 \
  --job_name=act-sourccey-demo-1 \
  --batch_size=8 \
  --steps=20000 \
  --save_freq=5000 \
  --wandb.enable=false \
  --policy.push_to_hub=false
03

Train X-VLA on a stronger GPU

If you have a stronger NVIDIA GPU with sufficient VRAM, we recommend fine-tuning the 0.9B-parameter X-VLA base model as the higher-capacity option. Install the X-VLA dependencies first:

Training computer · terminal
uv sync --locked --extra sourccey-desktop --extra training --extra xvla
Training computer · terminal
uv run lerobot-train \
  --dataset.repo_id=vulcan-studio/sourccey-demo-1 \
  --output_dir=outputs/train/xvla-sourccey-demo-1 \
  --job_name=xvla-sourccey-demo-1 \
  --policy.path=lerobot/xvla-base \
  --policy.dtype=bfloat16 \
  --policy.action_mode=auto \
  --policy.device=cuda \
  --policy.freeze_vision_encoder=false \
  --policy.freeze_language_encoder=false \
  --policy.train_policy_transformer=true \
  --policy.train_soft_prompts=true \
  --steps=20000 \
  --policy.push_to_hub=false

X-VLA starts from the pretrained lerobot/xvla-base checkpoint and uses bfloat16 to reduce memory use. If training runs out of memory, reduce the batch size before changing the model configuration.

04

Check the output

Training should complete without dataset feature or video-decoding errors. The ACT example creates this model directory:

Training computer · terminal
outputs/train/act-sourccey-demo-1/checkpoints/last/pretrained_model

About these examples: checks against a specific software version and a full hands-on walkthrough are still in progress. The source documentation below can help you and the team confirm the right instructions for your setup.

View upstream documentation·Read as Markdown
Vulcan Robotics / SourcceyBuild. Learn. Iterate.