Run Inference (Simulation)

Once you have a checkpoint, run the policy against the simulation.

Step 1 — Start the simulation

docker compose up

Step 2 — Run the inference node (in a new terminal)

docker compose run --rm smolvla

Options

Variable	Default	Description
`SMOLVLA_CHECKPOINT`	(empty — auto-discovers latest)	Override with a full path to pin a specific checkpoint, e.g. `/data/checkpoints/smolvla_ar4/checkpoints/040000/pretrained_model`
`CHECKPOINT_SEARCH_ROOT`	`/data/checkpoints`	Root directory to search for the latest checkpoint when `SMOLVLA_CHECKPOINT` is not set
`TASK`	`"place the teal cube on the top shelf"`	Task instruction sent to the model
`DEVICE`	`cuda`	`cpu` or `mps` to override; use `SMOLVLA_SERVER_URL` instead to offload to a remote GPU
`HZ`	`15.0`	Control frequency in Hz
`USE_SIM_TIME`	`true`	Must be `true` for sim, `false` for real hardware
`SPEED_SCALE`	`1.0`	Scale all joint motions (lower = more cautious)

# Auto-selects the latest checkpoint under ./data/checkpoints/
docker compose run --rm smolvla
 
# Pin a specific checkpoint step
SMOLVLA_CHECKPOINT=/data/checkpoints/smolvla_ar4/checkpoints/040000/pretrained_model \
  docker compose run --rm smolvla
 
# Different task
TASK="pick the red cube and place it on the shelf" docker compose run --rm smolvla

Step 1 — Start the simulation

Step 2 — Run the inference node (in a new terminal)

Options

On this page