Skip to main content
Experimental. OpenARM support is new and still being proven on hardware. Commands and setup may change without notice. Expect rough edges, stay at the e-stop, and tell us how it goes at team@dreamscalelabs.com. There is no shared model for the OpenARM: you run your own deployment, fine-tuned on demos recorded on your cell. The model runs in the cloud, so the Devbox is the only computer you need.

What you need

  • An Anvil OpenARM 2.0, one arm or two, with its Devbox and cameras.
  • Anvil’s control stack running on the Devbox with its inference config: set ARMS_CONTROL_CONFIG_FILE=config/openarm_v2_inference.yaml in .env.config, then start the stack as usual. In leader-follower mode the stack ignores joint commands, so a run refuses to start.
  • The Devbox online, on a wired connection if you can.
  • The e-stop within reach.
Dreamscale drives the arms and reads the cameras through Anvil’s stack. It never talks to the motors itself, so Anvil’s calibration and homing are untouched.

1. Install and set up the rig

On the Devbox:
Setup asks for the meta/info.json of the dataset your model was trained on and reads the layout from it: which arms, whether each arm’s gripper comes before or after its seven joints, and which cameras. Without one, it asks. It only reads the stack; nothing moves. Setup writes ~/.config/dreamscale/openarm.json. You can also write it yourself, for example from a coding agent:
  • url: the stack’s rosbridge address.
  • arms: the arms your model drives, in its order: left, right or both.
  • gripper: first if each arm’s eight values start with the gripper (what Anvil’s mcap-convert writes), last if they end with it.
  • cameras: up to three of waist, chest, wrist_l and wrist_r, the ones your model was trained on.

2. Check everything

It reads the stack without commanding anything: the joint states, each arm’s controller, the joint limits and the cameras, and it checks that your deployment is ready. Every line shows a ✓.

3. Your deployment

Record with Anvil’s web app and convert with mcap-convert, then fine-tune π0.5 and deploy it as on Deployments. The deployment needs:
  • cameras from observation.images.waist, observation.images.chest, observation.images.wrist_l and observation.images.wrist_r, at most three;
  • observation.state and action: 8 values for one arm or 16 for two. Each arm has seven joint positions in radians and its gripper’s opening in metres.
A run refuses a deployment whose cameras or width differ from the rig file’s.

4. Run it

The arms will move. Clear the workspace and keep the e-stop within reach. Software limits do not prevent collisions, and cloud inference does not replace your robot’s safety configuration.
  1. At Instruction, type a task.
  2. The run prints the model’s first prediction without moving anything. At Enter to start, press Enter.
  3. The arms run the task from their current pose. Press Ctrl+C to stop; you’re asked for the next instruction.
Every target stays inside the joint limits the stack describes. A command may lead the measured position by at most 0.1 rad and joints move at most 1 rad/s; lower them with --max-joint-step and --max-joint-velocity for a first run. When a run stops, the stack holds the arms where they are; stop the stack to park them.

Save a home pose (optional)

With the arms where every run should start:
dreamscale robots openarm home moves them back there slowly, and a run returns there when you stop it. Stuck? See Troubleshooting.