CLI reference - Dreamscale

Every public dreamscale command, generated from --help of version 0.1.0a258. Add --help (or -h) to any command to see the same list in your terminal. For step-by-step guides, start at Quick start.

What you need

The CLI: uv tool install dreamscale, or uv tool upgrade dreamscale to update.
An account for commands that reach the cloud: dreamscale login.

Commands

Command	What it does
`login`	Sign in on this machine.
`auth whoami`, `auth logout`	See and end this machine’s sign-in.
`doctor`	Check your setup.
`sim`	Run a task on a simulated benchmark.
`run so101`, `run yam`, `run franka`, `run ur`, `run xarm7`	Run tasks on a physical robot. Franka, Universal Robots and xArm 7 are experimental.
`robots so101 …`, `robots yam …`, `robots franka …`, `robots ur …`, `robots xarm7 …`	Set up and maintain a robot: setup, calibrate, home, release, recover.
`models list`	List the models your organization can run.
`sessions list`, `sessions stop`	See and stop your open sessions.
`deployments …`	Deploy and manage your fine-tuned models.
`train …`	Fine-tune a base model on your dataset, and manage runs.

Global options

Option	Description
`--version`	Print the installed version and exit. It also says when a newer release is out.
`--install-completion`	Install shell completion for your shell.
`--help`, `-h`	Show help for any command and exit.

Environment variables

Variable	Description
`DREAMSCALE_API_KEY`	API key to use instead of your saved login. When set, it wins over `dreamscale login`.
`HF_TOKEN`	Hugging Face token for private repos, used by `deployments` and `train` when `--hf-token` is not given.

Sign in on this machine. Opens your browser to sign in or create an account, then saves the login key to ~/.dreamscale/config.toml. Each machine holds one login key: logging in again revokes the key the last login saved. Login keys don’t appear on the API keys page, and a login never revokes a key you made there.

dreamscale login [OPTIONS]

Option	Default	Description
`--api-key TEXT`	—	Save this API key instead of signing in with a browser. Create one in Settings → API keys.

dreamscale auth

dreamscale auth login [OPTIONS]

Option	Default	Description
`--api-key TEXT`	—	Save this API key instead of signing in with a browser.

dreamscale auth whoami

Show the account and organization this machine is signed in as, and where the key came from: DREAMSCALE_API_KEY or your saved login.

dreamscale auth whoami

dreamscale auth logout

Sign out on this machine: revoke its login key and remove it from ~/.dreamscale/config.toml. A key you saved with dreamscale login --api-key is only removed from this machine. It keeps working until you revoke it in Settings → API keys. Logout leaves DREAMSCALE_API_KEY alone; unset DREAMSCALE_API_KEY to stop using it.

dreamscale auth logout

dreamscale doctor

Check your setup: the CLI, your account and, for a robot, its hardware and your access to the model run would use. Every check prints a ✓ or a fix.

dreamscale doctor [MODE] [OPTIONS]

Argument or option	Default	Description
`MODE`	`core`	What to check: `core`, `sim`, `all`, `so101`, `yam`, or `franka`, `ur` or `xarm7` (experimental).
`--config PATH`	saved rig config	YAM, Franka, UR and xArm 7 only: the rig config to check.
`--model TEXT`	—	With a robot: the model to check access to. Without it, `doctor` checks the robot only.

dreamscale doctor so101

dreamscale sim

Run a task on a simulated benchmark. The simulator runs on your computer and the model runs in the cloud. The first run of a benchmark downloads it.

dreamscale sim BENCHMARK --model MODEL [OPTIONS]

Argument or option	Default	Description
`BENCHMARK`	—	`libero`, `robocasa365`, `molmospaces` or `roboeval`.
`--model TEXT`	required	Model to run, such as `pi0.5-libero` for `libero`. `dreamscale models list` lists yours.
`--task-suite TEXT`	per benchmark	`libero`: `libero_spatial` (default), `libero_object`, `libero_goal`, `libero_10` or `libero_90`. `robocasa365`: `atomic` or `composite`. `molmospaces`: a benchmark or `all` (default). `roboeval`: a task (default `lift_pot`).
`--task-id TEXT`	`0`	Task within the suite. `molmospaces` also takes `all`.
`--task-seed INTEGER`	`0`	Seed for the task’s initial state.
`--max-steps INTEGER`	the task’s horizon	Action-step limit, at least 1.
`--max-duration FLOAT`	—	Time limit in seconds (at least 0.1), instead of `--max-steps`.
`--non-interactive`	off	No viewer window; print a JSON summary at the end.
`--transport`	`auto`	`auto` (QUIC, then relay), `quic` or `relay`.
`--region TEXT`	—	Where to serve: `nearest`, `available` or `us-west-2`.
`--output-root PATH`	—	Save the run’s artifacts (video, timing) to this directory.
`--verbose`, `-v`	off	Show diagnostic details.

dreamscale sim libero --model pi0.5-libero --task-suite libero_object --task-id 3

dreamscale run

Run tasks on a physical robot. Each task homes the arm, previews the first plan without moving, runs on Enter until Ctrl+C, then returns home and asks for the next instruction. See SO-101, Bimanual YAM or Franka / DROID (experimental) for the full walkthrough.

dreamscale run moves the robot. Enforce joint, speed and workspace limits locally, and keep a physical e-stop within reach. Cloud inference does not replace your robot’s safety controller.

dreamscale run so101

Run tasks on an SO-101. Joint limits are in degrees.

dreamscale run so101 --model MODEL [OPTIONS]

Option	Default	Description
`--model TEXT`	required	A model from `dreamscale models list`, such as `molmoact2-so101`, `flux-3-action-so101` or `pi0.5-so101`, or a deployment as `<account>/<slug>`. Optional when `--deployment` names a deployment, which runs its own model.
`--instruction TEXT`	prompt for tasks	Run this one task, then exit.
`--mode TEXT`	`async`	`async` (replan continuously) or `sync` (finish each chunk before the next).
`--max-actions INTEGER`	run until Ctrl+C	Action limit per task, 1–18000.
`--region TEXT`	`available`	Where to serve: `nearest`, `available` or `us-west-2`.
`--provider TEXT`	—	Advanced: `aws` or `modal`.
`--speed TEXT`	`standard`	`standard`, `fast` or `ultrafast`, where the model supports it.
`--rtc TEXT`	`auto`	Advanced real-time chunking: `auto` (the model’s default), `on`, `off`, `train_time` or `inference_time`.
`--deployment TEXT`	—	Deployment to run: `<account>/<slug>`, its slug, its `ckpt_…` ID, or `base`.
`--context TEXT`	`auto`	Frame history, for deployments that use it: `auto` (the deployment’s own), `on` (last 3 frames) or `off` (1 frame).
`--effort TEXT`	`medium`	Reasoning effort, for models that take one: `low`, `medium`, `high`, `xhigh` or `max`.
`--max-cost FLOAT`	none	Credit limit per task, for models billed by usage.
`--max-joint-step FLOAT`	`45`	Most degrees a command may lead the measured joint position.
`--max-joint-velocity FLOAT`	`150`	Joint speed cap in degrees per second, homing included.
`--action-smoothing TEXT`	`auto`	Chunk blending for `async` with RTC off: `auto`, `motion` or `output`.
`--trace-dir PATH`	—	Save a recording of each task under this directory.
`--trace-images`	off	Include camera frames in recordings (needs `--trace-dir`).
`--verbose`	off	Show diagnostic details.

dreamscale run so101 --model pi0.5-so101 --instruction "pick up the red block"

dreamscale run yam

Run tasks on a bimanual YAM. Takes the same options as run so101, plus --config. Joint limits are in radians.

dreamscale run yam --model MODEL [OPTIONS]

Option	Default	Description
`--model TEXT`	required	A model from `dreamscale models list`, such as `molmoact2-bimanual-yam` or `pi0.5-bimanualyam`, or a deployment as `<account>/<slug>`. Optional when `--deployment` names a deployment, which runs its own model.
`--config PATH`	saved rig config	The rig config from `dreamscale robots yam setup`.
`--instruction TEXT`	prompt for tasks	Run this one task, then exit.
`--mode TEXT`	`async`	`async` (replan continuously) or `sync` (finish each chunk before the next).
`--max-actions INTEGER`	run until Ctrl+C	Action limit per task, 1–18000.
`--region TEXT`	`available`	Where to serve: `nearest`, `available` or `us-west-2`.
`--provider TEXT`	—	Advanced: `aws` or `modal`.
`--speed TEXT`	`standard`	`standard`, `fast` or `ultrafast`, where the model supports it.
`--rtc TEXT`	`auto`	Advanced real-time chunking: `auto` (the model’s default), `on`, `off`, `train_time` or `inference_time`.
`--deployment TEXT`	—	Deployment to run: `<account>/<slug>`, its slug, its `ckpt_…` ID, or `base`.
`--context TEXT`	`auto`	Frame history, for deployments that use it: `auto` (the deployment’s own), `on` (last 3 frames) or `off` (1 frame).
`--effort TEXT`	`medium`	Reasoning effort, for models that take one: `low`, `medium`, `high`, `xhigh` or `max`.
`--max-cost FLOAT`	none	Credit limit per task, for models billed by usage.
`--max-joint-step FLOAT`	`0.15`	Most radians a command may lead the measured joint position.
`--max-joint-velocity FLOAT`	`1.0`	Joint speed cap in radians per second, homing included.
`--action-smoothing TEXT`	`auto`	Chunk blending for `async` with RTC off: `auto`, `motion` or `output`.
`--trace-dir PATH`	—	Save a recording of each task under this directory.
`--trace-images`	off	Include camera frames in recordings (needs `--trace-dir`).
`--verbose`	off	Show diagnostic details.

dreamscale run yam --model acme-robotics/towel-fold --instruction "fold the towel"

dreamscale run franka

Run tasks on a Franka arm. Experimental: see Franka / DROID. Takes the same options as run yam. Joint values are in radians, and the speed and step limits are off unless you set them: the arm’s own controller tracks the model.

dreamscale run franka --model MODEL [OPTIONS]

Option	Default	Description
`--model TEXT`	required	A model from `dreamscale models list`, such as `molmoact2-droid`, or a deployment as `<account>/<slug>`.
`--config PATH`	saved rig config	The rig config from `dreamscale robots franka setup`.
`--instruction TEXT`	prompt for tasks	Run this one task, then exit.
`--mode TEXT`	`async`	`async` (replan continuously) or `sync` (finish each chunk before the next).
`--max-actions INTEGER`	run until Ctrl+C	Action limit per task, 1–18000.
`--region TEXT`	`available`	Where to serve: `nearest`, `available` or `us-west-2`.
`--provider TEXT`	—	Advanced: `aws` or `modal`.
`--speed TEXT`	`standard`	`standard`, `fast` or `ultrafast`, where the model supports it.
`--rtc TEXT`	`auto`	Advanced real-time chunking: `auto` (the model’s default), `on`, `off`, `train_time` or `inference_time`.
`--deployment TEXT`	—	Deployment to run: `<account>/<slug>`, its slug, its `ckpt_…` ID, or `base`.
`--context TEXT`	`auto`	Frame history, for deployments that use it: `auto` (the deployment’s own), `on` (last 3 frames) or `off` (1 frame).
`--effort TEXT`	`medium`	Reasoning effort, for models that take one: `low`, `medium`, `high`, `xhigh` or `max`.
`--max-cost FLOAT`	none	Credit limit per task, for models billed by usage.
`--max-joint-step FLOAT`	none	Most radians a command may lead the measured joint position.
`--max-joint-velocity FLOAT`	none	Joint speed cap in radians per second, homing included.
`--action-smoothing TEXT`	`auto`	Chunk blending for `async` with RTC off: `auto`, `motion` or `output`.
`--trace-dir PATH`	—	Save a recording of each task under this directory.
`--trace-images`	off	Include camera frames in recordings (needs `--trace-dir`).
`--verbose`	off	Show diagnostic details.

dreamscale run franka --model molmoact2-droid --instruction "pick up the marker"

dreamscale run ur

Run your own π0.5 deployment on a Universal Robots arm. Experimental: see Universal Robots. Takes the same options as run yam. Joint limits are in radians; on arms longer than a UR7e the speed and step caps shrink with reach.

dreamscale run ur --model DEPLOYMENT [OPTIONS]

Option	Default	Description
`--model TEXT`	required	Your π0.5 deployment for this rig, as `<account>/<slug>`.
`--config PATH`	saved rig config	The rig config from `dreamscale robots ur setup`.
`--instruction TEXT`	prompt for tasks	Run this one task, then exit.
`--mode TEXT`	`async`	`async` (replan continuously) or `sync` (finish each chunk before the next).
`--max-actions INTEGER`	run until Ctrl+C	Action limit per task, 1–18000.
`--region TEXT`	`available`	Where to serve: `nearest`, `available` or `us-west-2`.
`--provider TEXT`	—	Advanced: `aws` or `modal`.
`--speed TEXT`	`standard`	`standard`, `fast` or `ultrafast`, where the model supports it.
`--rtc TEXT`	`auto`	Advanced real-time chunking: `auto` (the model’s default), `on`, `off`, `train_time` or `inference_time`.
`--deployment TEXT`	—	Deployment to run: `<account>/<slug>`, its slug, its `ckpt_…` ID, or `base`.
`--context TEXT`	`auto`	Frame history, for deployments that use it: `auto` (the deployment’s own), `on` (last 3 frames) or `off` (1 frame).
`--effort TEXT`	`medium`	Reasoning effort, for models that take one: `low`, `medium`, `high`, `xhigh` or `max`.
`--max-cost FLOAT`	none	Credit limit per task, for models billed by usage.
`--max-joint-step FLOAT`	`0.3`	Most radians a command may lead the measured joint position.
`--max-joint-velocity FLOAT`	`1.0`	Joint speed cap in radians per second, homing included.
`--action-smoothing TEXT`	`auto`	Chunk blending for `async` with RTC off: `auto`, `motion` or `output`.
`--trace-dir PATH`	—	Save a recording of each task under this directory.
`--trace-images`	off	Include camera frames in recordings (needs `--trace-dir`).
`--verbose`	off	Show diagnostic details.

dreamscale run ur --model acme-robotics/cup-to-plate --instruction "put the cup on the plate"

dreamscale run xarm7

Run your own deployment on two xArm 7s. Experimental: see Two xArm 7s. Takes the same options as run yam.

dreamscale run xarm7 --model DEPLOYMENT [OPTIONS]

Option	Default	Description
`--model TEXT`	required	Your deployment for this rig, as `<account>/<slug>`.
`--config PATH`	saved rig config	The rig config from `dreamscale robots xarm7 setup`.
`--instruction TEXT`	prompt for tasks	Run this one task, then exit.
`--mode TEXT`	`async`	`async` (replan continuously) or `sync` (finish each chunk before the next).
`--max-actions INTEGER`	run until Ctrl+C	Action limit per task, 1–18000.
`--region TEXT`	`available`	Where to serve: `nearest`, `available` or `us-west-2`.
`--provider TEXT`	—	Advanced: `aws` or `modal`.
`--speed TEXT`	`standard`	`standard`, `fast` or `ultrafast`, where the model supports it.
`--rtc TEXT`	`auto`	Advanced real-time chunking: `auto` (the model’s default), `on`, `off`, `train_time` or `inference_time`.
`--deployment TEXT`	—	Deployment to run: `<account>/<slug>`, its slug, its `ckpt_…` ID, or `base`.
`--context TEXT`	`auto`	Frame history, for deployments that use it: `auto` (the deployment’s own), `on` (last 3 frames) or `off` (1 frame).
`--effort TEXT`	`medium`	Reasoning effort, for models that take one: `low`, `medium`, `high`, `xhigh` or `max`.
`--max-cost FLOAT`	none	Credit limit per task, for models billed by usage.
`--max-joint-step FLOAT`	`0.2`	Most radians a command may lead the measured joint position.
`--max-joint-velocity FLOAT`	`0.5`	Joint speed cap in radians per second, homing included.
`--action-smoothing TEXT`	`auto`	Chunk blending for `async` with RTC off: `auto`, `motion` or `output`.
`--trace-dir PATH`	—	Save a recording of each task under this directory.
`--trace-images`	off	Include camera frames in recordings (needs `--trace-dir`).
`--verbose`	off	Show diagnostic details.

dreamscale run xarm7 --model acme-robotics/refill --instruction "pick up the funnel from the table"

dreamscale robots

Set up and maintain robots. Setup installs the robot’s support the first time it runs.

dreamscale robots so101 setup

Install SO-101 support, then choose its cameras and serial port on a setup page in your browser.

dreamscale robots so101 setup [OPTIONS]

Option	Default	Description
`--cli`	off	Answer terminal prompts instead of using the setup page.
`--host TEXT`	`0.0.0.0`	Setup page server host.
`--port INTEGER`	`8765`	Setup page server port, 1–65535.

dreamscale robots so101 calibrate

Calibrate the arm, step by step. Takes no options.

dreamscale robots so101 calibrate

dreamscale robots so101 home

Move to the home pose, then release motor torque. run so101 starts and ends each task at this pose.

dreamscale robots so101 home [OPTIONS]

Option	Default	Description
`--set`	off	Save the current pose as home. The motors go limp so you can pose the arm.

dreamscale robots so101 release

Release motor torque. Support the arm first. Takes no options.

dreamscale robots so101 release

dreamscale robots yam setup

Install YAM support and assign cameras and CAN ports, with live camera previews over your LAN.

dreamscale robots yam setup [OPTIONS]

Option	Default	Description
`--cli`	off	Pick from numbered terminal prompts instead of LAN camera previews.
`--config PATH`	saved rig config	Rig config path to write.
`--host TEXT`	`0.0.0.0`	Camera preview server host.
`--port INTEGER`	`8765`	Camera preview server port, 1–65535.

dreamscale robots yam home

Move to the home pose, then release motor torque.

dreamscale robots yam home [OPTIONS]

Option	Default	Description
`--config PATH`	saved rig config	Rig config path.

dreamscale robots yam release

Release motor torque. Support the arms first.

dreamscale robots yam release [OPTIONS]

Option	Default	Description
`--config PATH`	saved rig config	Rig config path.

dreamscale robots franka setup

Install Franka support, check that it can reach the arm, and choose its cameras on a setup page in your browser. Experimental: see Franka / DROID.

dreamscale robots franka setup [OPTIONS]

Option	Default	Description
`--cli`	off	Answer terminal prompts instead of using the setup page.
`--config PATH`	saved rig config	Rig config path to write.
`--host TEXT`	`0.0.0.0`	Setup page server host.
`--port INTEGER`	`8765`	Setup page server port, 1–65535.

dreamscale robots franka recover

Clear the arm’s error after a reflex, once you have checked why it stopped.

dreamscale robots franka recover [OPTIONS]

Option	Default	Description
`--config PATH`	saved rig config	Rig config path.

dreamscale robots ur setup

Experimental. Install Universal Robots support, then choose the arm model, control box address, gripper and cameras on a setup page in your browser. Nothing moves.

dreamscale robots ur setup [OPTIONS]

Option	Default	Description
`--cli`	off	Answer terminal prompts instead of using the setup page.
`--config PATH`	saved rig config	Rig config path to write.
`--host TEXT`	`0.0.0.0`	Setup page server host.
`--port INTEGER`	`8765`	Setup page server port, 1–65535.

dreamscale robots ur home

Experimental. Move slowly to the home pose and open the gripper; the arm then holds.

dreamscale robots ur home [OPTIONS]

Option	Default	Description
`--set`	off	Save where the arm is now as home. Move it there with freedrive first.
`--config PATH`	saved rig config	Rig config path.

dreamscale robots xarm7 setup

Experimental. Install xArm 7 support, then enter each arm’s control box address and gripper and assign the three cameras, in the terminal. Nothing moves.

dreamscale robots xarm7 setup [OPTIONS]

Option	Default	Description
`--config PATH`	saved rig config	Rig config path to write.

dreamscale robots xarm7 home

Experimental. Open the grippers and move both arms slowly to the home pose; the arms then hold.

dreamscale robots xarm7 home [OPTIONS]

Option	Default	Description
`--set`	off	Save where the arms are now as home. Move them there in UFACTORY Studio first.
`--config PATH`	saved rig config	Rig config path.

dreamscale models

dreamscale models list

List the models your organization can run, with the robot each runs on and its state (Ready, Busy, Starting or Cold start).

dreamscale models list [OPTIONS]

Option	Default	Description
`--json`	off	Print raw JSON.
`--verbose`, `-v`	off	Show where each model serves and its live state.

dreamscale sessions

A session is one connection to a model. Billing runs while it is open.

dreamscale sessions list

List your open sessions.

dreamscale sessions list [OPTIONS]

Option	Default	Description
`--json`	off	Print raw JSON.

dreamscale sessions stop

Stop a session, or all of them, to end billing.

dreamscale sessions stop [SESSION_ID] [OPTIONS]

Argument or option	Default	Description
`SESSION_ID`	—	The session to stop, from `dreamscale sessions list`.
`--all`	off	Stop all your open sessions.

dreamscale deployments

Deploy and manage your fine-tuned models. Wherever a command takes a DEPLOYMENT, give its <account>/<slug> name, its bare slug or its ckpt_… ID. See Deployments for the walkthrough.

dreamscale deployments create

Deploy fine-tuned weights of π0.5 or FLUX 3 Action (pick with --base) from Hugging Face (any huggingface.co link works, including a checkpoint folder’s page) or from a local folder. From a folder, only the files the build reads are uploaded, and a re-run sends nothing new. Values not given are detected from the checkpoint; run inspect to see them.

dreamscale deployments create REPO [OPTIONS]

Argument or option	Default	Description
`REPO`	—	Hugging Face repo (`hf://owner/name`, `owner/name` or a `huggingface.co` link), or a local π0.5 checkpoint folder (`./folder`).
`--base TEXT`	`pi0.5`	Base model: `pi0.5` or `flux-3-action-base`. A folder `dreamscale train download` made needs none: its base is read from it.
`--preset TEXT`	detected	Robot shape: `bimanual_yam`, `so101`, `arx5` or `blank`.
`--camera TEXT`	detected	`key=slot` (slots: `base`, `left_wrist`, `right_wrist`). Repeat it, in the order your robot sends cameras.
`--hz FLOAT`	detected	Control rate in Hz the weights expect.
`--path TEXT`	the latest	π0.5: the checkpoint in the repo, for example `checkpoints/55000`.
`--horizon INTEGER`	`50`	π0.5: actions the model predicts per chunk.
`--executed INTEGER`	the horizon	π0.5: actions executed per chunk.
`--token-length INTEGER`	`200`	π0.5: prompt length in tokens.
`--no-state-in-prompt`	off	π0.5: the checkpoint was trained without the state in its prompt.
`--delta-mask TEXT`	detected	π0.5: `1` or `0` per action, comma separated (`1` = relative to the state), or `none`.
`--rtc TEXT`	detected	π0.5: `off`, `test-time` or `trained`.
`--rtc-max-delay INTEGER`	—	π0.5: the delay trained RTC was trained with, 1–15.
`--rtc-prefix INTEGER`	`4`	Trained RTC: at most this many planned actions pinned per request, within what the checkpoint was trained on.
`--revision TEXT`	default branch	Branch, tag or commit.
`--hf-token TEXT`	`HF_TOKEN`, else your connected token	Hugging Face token for this request.
`--region TEXT`	`auto`	Where it serves: a region id (`us-west`, `us-central`, `us-east`, `uk`, `eu-west`, `japan`, `india`, `australia`), or `auto`: the nearest one its base serves, measured from this machine.
`--keep-warm TEXT`	`off`	How long its server stays up after the last robot leaves: `off`, `10m`, `30m`, `1h`, `2h`, `6h`, `12h` or `always`. Billed as server time.
`--slug TEXT`	from the repo and step	Its name: the `<model>` part of `<account>/<model>`.
`--region-fallback / --no-region-fallback`	on	If its region is full, start its server in the next-closest region (a little farther).
`--no-wait`	off	Return once the build starts.

dreamscale deployments create hf://your-org/towel-fold --preset bimanual_yam --slug towel-fold

dreamscale deployments inspect

Show a repo’s (or a local folder’s) robot shape, what was detected and where from. Nothing is uploaded.

dreamscale deployments inspect REPO [OPTIONS]

Argument or option	Default	Description
`REPO`	—	Hugging Face repo (`hf://owner/name`, `owner/name` or a `huggingface.co` link), or a local π0.5 checkpoint folder.
`--base TEXT`	`auto`	`pi0.5`, `flux-3-action-base` or `auto`.
`--path TEXT`	the latest	π0.5: the checkpoint in the repo, for example `checkpoints/55000`.
`--revision TEXT`	default branch	Branch, tag or commit.
`--hf-token TEXT`	`HF_TOKEN`, else your connected token	Hugging Face token for this request.
`--json`	off	Print raw JSON.

dreamscale deployments list

List your organization’s deployments.

dreamscale deployments list [OPTIONS]

Option	Default	Description
`--model TEXT`	all	Only deployments of this base model.
`--json`	off	Print raw JSON.

dreamscale deployments show

Show a deployment’s status, build progress or error.

dreamscale deployments show DEPLOYMENT [OPTIONS]

Argument or option	Default	Description
`DEPLOYMENT`	—	The deployment’s name or ID.
`--json`	off	Print the deployment as JSON. Its `status` is `ready` once it can run.

dreamscale deployments keep-warm

Keep a deployment’s server warm between runs. Warm time is billed as server time.

dreamscale deployments keep-warm DEPLOYMENT VALUE [OPTIONS]

Argument or option	Default	Description
`VALUE`	—	How long its server stays up after the last robot leaves: `off`, `10m`, `30m`, `1h`, `2h`, `6h`, `12h` or `always`.
`--take-over-from TEXT`	—	Move always-on from this deployment once this one is warm.
`--robots INTEGER`	—	How many robots stay ready, 1–64. Billed as at least that many robots while kept warm. With one robot running, `2` keeps a spare: if its server stops, the robot moves to the spare in about 5 seconds.
`--now`	off	Start it now rather than at the first robot, kept for `VALUE` from now, and wait until it is ready. Ctrl-C stops waiting, not the server.

dreamscale deployments keep-warm towel-fold 30m --now

dreamscale deployments rtc

Set a π0.5 or FLUX.3 Action deployment’s real-time chunking for its new sessions. No rebuild. A FLUX.3 Action deployment of a run trained with --rtc runs it by default.

dreamscale deployments rtc DEPLOYMENT METHOD [OPTIONS]

Argument or option	Default	Description
`METHOD`	—	`off`, `test-time` or `trained`. FLUX.3 Action: `off` or `trained`.
`--max-delay INTEGER`	—	π0.5: the delay trained RTC was trained with, 1–15. FLUX.3 Action: the most committed actions kept per request; by default all its run trained.

dreamscale deployments fallback

Turn region fallback on or off for a deployment’s next server. No rebuild. On (the default), if its region is full, its server starts in the next-closest region (a little farther). Off, a robot waits for its own region.

dreamscale deployments fallback DEPLOYMENT SETTING

Argument or option	Default	Description
`DEPLOYMENT`	—	The deployment’s name or ID.
`SETTING`	—	`on` or `off`.

dreamscale deployments rename

Rename a deployment. Its name is its slug. The old name keeps working until another of your deployments takes it.

dreamscale deployments rename DEPLOYMENT [SLUG]

Argument	Default	Description
`SLUG`	—	Its new slug: the `<model>` part of `<account>/<model>`.

dreamscale deployments restart

Stop the deployment’s server now. With Always on, a fresh one starts at once. A connected robot loses its connection and reconnects to a new server on its next request; the SDK holds it still meanwhile.

dreamscale deployments restart DEPLOYMENT

dreamscale deployments stop

Stop the deployment’s server now and keep none warm (turns its keep-warm off). The next robot starts a fresh server; a connected robot loses its connection.

dreamscale deployments stop DEPLOYMENT

dreamscale deployments rm

Permanently delete a deployment. Only organization owners can delete, and it cannot be undone. It asks first; robots running it are disconnected. With no terminal to ask in (a script or CI), pass --yes.

dreamscale deployments rm DEPLOYMENT [OPTIONS]

Argument or option	Default	Description
`DEPLOYMENT`	—	The deployment’s name or ID.
`--yes`, `-y`	off	Delete without asking.

dreamscale datasets

Keep your LeRobot datasets on Dreamscale and train on them. See Datasets.

dreamscale datasets upload

Upload a folder: what’s new is sent, and a new version recorded. Rerun it to resume.

dreamscale datasets upload FOLDER [OPTIONS]

Argument or option	Default	Description
`FOLDER`	—	The LeRobot dataset’s root folder.
`--name TEXT`	the folder’s	The dataset’s name.
`--no-check`	off	Skip reading the recordings here (training checks them again).
`--json`	off	Print the dataset as JSON.

dreamscale datasets import

Copy a Hugging Face dataset in, on our side. Importing again copies only what changed.

dreamscale datasets import SOURCE [OPTIONS]

Argument or option	Default	Description
`SOURCE`	—	`hf://owner/name`, with `/folder` and `@revision` if you need them.
`--name TEXT`	the repo’s	The dataset’s name.
`--hf-token TEXT`	`HF_TOKEN`, else your connected token	Hugging Face token for this request.
`--no-wait`	off	Return once the copy starts.
`--json`	off	Print the dataset as JSON.

dreamscale datasets list

Your datasets.

dreamscale datasets list [OPTIONS]

Argument or option	Default	Description
`--json`	off	Print the datasets as JSON.

dreamscale datasets show

A dataset’s summary, versions and runs; follows a running import.

dreamscale datasets show NAME [OPTIONS]

Argument or option	Default	Description
`NAME`	—	`NAME`, or `NAME@N` for one version.
`--json`	off	Print the dataset as JSON.

dreamscale datasets download

Download a version into a LeRobot folder. Rerun it to resume.

dreamscale datasets download NAME [FOLDER]

Argument or option	Default	Description
`NAME`	—	`NAME`, or `NAME@N` for one version.
`FOLDER`	`./NAME`	Where to.

dreamscale datasets rename

Rename a dataset.

dreamscale datasets rename NAME NEW_NAME

Argument or option	Default	Description
`NAME`	—	The dataset.
`NEW_NAME`	—	Its new name (the old one keeps working until another dataset takes it).

dreamscale datasets rm

Delete a dataset or one version, for good. What was made from it keeps working.

dreamscale datasets rm NAME [OPTIONS]

Argument or option	Default	Description
`NAME`	—	`NAME` for the whole dataset, or `NAME@N` for one version.
`--yes`, `-y`	off	Delete without asking.

dreamscale train

Fine-tune a base model on your LeRobot dataset, on Dreamscale’s GPUs. Name a dataset and a base to start a run: the CLI quotes the price and time, and starts once you confirm. See Training for the walkthrough.

dreamscale train --dataset DATASET --base BASE [OPTIONS]

Argument or option	Default	Description
`--dataset TEXT`	required	One of your datasets (`towels`, `towels@2`, `acme/towels`) or a Hugging Face one (`hf://owner/name`). Repeat it to train on several Hugging Face datasets together: they must match after the column mapping.
`--base TEXT`	required	Start from `pi0.5` (π0.5 base), `flux-3-action-base` (FLUX.3 Action), `pi0.5-rtc-yam` (π0.5 RTC YAM, trained on bimanual YAM, for any robot, where your organization has it), one of your runs (`tr_…`): its latest checkpoint, or one of your deployments (its name or ID): the checkpoint it serves, including a π0.5 or FLUX.3 Action checkpoint you imported.
`--base-step INTEGER`	the latest	With `--base tr_…`: start from this step’s checkpoint.
`--steps INTEGER`	auto	Training length in steps, 10–60000.
`--max-usd FLOAT`	—	Train as long as this many dollars buys, instead of `--steps`. At least 0.01.
`--batch-size INTEGER`	from the data	Global batch: π0.5 64, 128, 256 or 512; FLUX.3 Action 32, 64, 128 or 256.
`--learning-rate FLOAT`	the recipe’s	Peak learning rate, at most 0.01. Continuing from one of your runs, the default is half the recipe’s, at most that run’s.
`--warmup-steps INTEGER`	the recipe’s	Steps rising to the peak learning rate.
`--schedule TEXT`	the recipe’s	π0.5 only: `cosine` (decays over the run) or `constant` after the warm-up.
`--checkpoint-every INTEGER`	5 a run	Save a deployable checkpoint every N steps, at most 20 a run. The last step always saves.
`--eval-fraction FLOAT`	off	Hold out this share of episodes (0.05–0.3) and report a held-out loss per checkpoint. Held-out loss is a poor proxy for task success in diffusion and flow-matching policies: try checkpoints on the robot.
`--quote`	off	Show the quote and stop: nothing is started, even with `--yes`.
`--rtc`	the base’s	Train with real-time chunking. By default a run trains what its base does: on for π0.5, off for FLUX.3 Action, where it is experimental. The model also runs without committed actions.
`--no-rtc`	the base’s	Train without real-time chunking.
`--name TEXT`	the dataset	Display name.
`--hf-token TEXT`	`HF_TOKEN`, else your connected token	Hugging Face token for this request.
`--yes`, `-y`	off	Accept the quote without asking.
`--no-wait`	off	Return once the run starts.
`--json`	off	Print the run (with `--quote`, the quote) as JSON.
`--no-clean`	off	Train every episode as recorded. By default, auto-clean cuts out faulty stretches, trims idle starts and ends, and drops episodes that can’t be trained on.
`--state TEXT`	`observation.state`	The state’s columns, joined in order (`a+b`).
`--action TEXT`	`action`	The action’s columns, joined in order (`a+b`).
`--sample-weight COLUMN`	—	A one-value column with each episode’s weight: its samples are drawn that much more often; 0 leaves it out. π0.5 and FLUX.3 Action.
`--loss-mask COLUMN`	—	A one-value column, 1 on frames whose actions shouldn’t train (they stay as context). π0.5 and FLUX.3 Action.
`--camera TEXT`	by camera name	A camera and its image slot, for example `top=base` (slots: `base`, `left_wrist`, `right_wrist`). Repeat it for each camera.

dreamscale train --dataset towel-fold --base pi0.5 --max-usd 50

dreamscale train list

List your organization’s training runs.

dreamscale train list [OPTIONS]

Option	Default	Description
`--json`	off	Print raw JSON.

dreamscale train show

Show a training run. While it runs, it follows it live; Ctrl+C stops watching, not training.

dreamscale train show RUN_ID [OPTIONS]

Argument or option	Default	Description
`RUN_ID`	—	Training run ID (`tr_…`).
`--json`	off	Print raw JSON.

dreamscale train stop

Stop a training run. You pay for the steps it completed, and it saves a checkpoint you can deploy. Resume it within 30 days.

dreamscale train stop RUN_ID [OPTIONS]

Argument or option	Default	Description
`RUN_ID`	—	Training run ID (`tr_…`).
`--yes`, `-y`	off	Stop without asking.

dreamscale train resume

Continue a stopped training run for its remaining steps, at the same rate. It shows the price first. The run restores its saved state before it trains again, usually within minutes; while it waits for the cloud to start a worker for that, the run says how long it has waited.

dreamscale train resume RUN_ID [OPTIONS]

Argument or option	Default	Description
`RUN_ID`	—	Training run ID (`tr_…`).
`--yes`, `-y`	off	Accept the price without asking.
`--no-wait`	off	Return once it resumes.
`--json`	off	Print the run as JSON.

dreamscale train rm

Permanently delete a training run. One still going is canceled first. Deployments made from its checkpoints keep running.

dreamscale train rm RUN_ID [OPTIONS]

Argument or option	Default	Description
`RUN_ID`	—	Training run ID (`tr_…`).
`--yes`, `-y`	off	Delete without asking.

dreamscale train download

Download a checkpoint and its training contract as a tar archive.

dreamscale train download RUN_ID [OPTIONS]

Argument or option	Default	Description
`RUN_ID`	—	Training run ID (`tr_…`).
`--step INTEGER`	the latest	Checkpoint step.
`--output`, `-o TEXT`	here	File or folder to save it in.
`--link`	off	Print a signed link, valid for an hour, instead of downloading.

dreamscale train deploy

Deploy a checkpoint of a training run. No robot details are needed: the dataset defined the robot. It prints the command that runs it.

dreamscale train deploy RUN_ID [OPTIONS]

Argument or option	Default	Description
`RUN_ID`	—	Training run ID (`tr_…`).
`--step INTEGER`	the latest	Checkpoint step.
`--region TEXT`	`auto`	Where it serves: a region id such as `us-west` or `japan`, or `auto`: the nearest one the run’s base serves, measured from this machine.
`--no-wait`	off	Return once the build starts.
`--region-fallback / --no-region-fallback`	on	If its region is full, start its server in the next-closest region (a little farther).

Stuck? See Troubleshooting or email team@dreamscalelabs.com.

​What you need

​Commands

​Global options

​Environment variables

​dreamscale login

​dreamscale auth

​dreamscale auth login

​dreamscale auth whoami

​dreamscale auth logout

​dreamscale doctor

​dreamscale sim

​dreamscale run

​dreamscale run so101

​dreamscale run yam

​dreamscale run franka

​dreamscale run ur

​dreamscale run xarm7

​dreamscale robots

​dreamscale robots so101 setup

​dreamscale robots so101 calibrate

​dreamscale robots so101 home

​dreamscale robots so101 release

​dreamscale robots yam setup

​dreamscale robots yam home

​dreamscale robots yam release

​dreamscale robots franka setup

​dreamscale robots franka recover

​dreamscale robots ur setup

​dreamscale robots ur home

​dreamscale robots xarm7 setup

​dreamscale robots xarm7 home

​dreamscale models

​dreamscale models list

​dreamscale sessions

​dreamscale sessions list

​dreamscale sessions stop

​dreamscale deployments

​dreamscale deployments create

​dreamscale deployments inspect

​dreamscale deployments list

​dreamscale deployments show

​dreamscale deployments keep-warm

​dreamscale deployments rtc

​dreamscale deployments fallback

​dreamscale deployments rename

​dreamscale deployments restart

​dreamscale deployments stop

​dreamscale deployments rm

​dreamscale datasets

​dreamscale datasets upload

​dreamscale datasets import

​dreamscale datasets list

​dreamscale datasets show

​dreamscale datasets download

​dreamscale datasets rename

​dreamscale datasets rm

​dreamscale train

​dreamscale train list

​dreamscale train show

​dreamscale train stop

​dreamscale train resume

​dreamscale train rm

​dreamscale train download

​dreamscale train deploy

What you need

Commands

Global options

Environment variables

dreamscale login

dreamscale auth

dreamscale auth login

dreamscale auth whoami

dreamscale auth logout

dreamscale doctor

dreamscale sim

dreamscale run

dreamscale run so101

dreamscale run yam

dreamscale run franka

dreamscale run ur

dreamscale run xarm7

dreamscale robots

dreamscale robots so101 setup

dreamscale robots so101 calibrate

dreamscale robots so101 home

dreamscale robots so101 release

dreamscale robots yam setup

dreamscale robots yam home

dreamscale robots yam release

dreamscale robots franka setup

dreamscale robots franka recover

dreamscale robots ur setup

dreamscale robots ur home

dreamscale robots xarm7 setup

dreamscale robots xarm7 home

dreamscale models

dreamscale models list

dreamscale sessions

dreamscale sessions list

dreamscale sessions stop

dreamscale deployments

dreamscale deployments create

dreamscale deployments inspect

dreamscale deployments list

dreamscale deployments show

dreamscale deployments keep-warm

dreamscale deployments rtc

dreamscale deployments fallback

dreamscale deployments rename

dreamscale deployments restart

dreamscale deployments stop

dreamscale deployments rm

dreamscale datasets

dreamscale datasets upload

dreamscale datasets import

dreamscale datasets list

dreamscale datasets show

dreamscale datasets download

dreamscale datasets rename

dreamscale datasets rm

dreamscale train

dreamscale train list

dreamscale train show

dreamscale train stop

dreamscale train resume

dreamscale train rm

dreamscale train download

dreamscale train deploy