API Reference

VLAModel

The main model class for loading and running VLA models.

`VLAModel.from_preset(name)`

Load a model by preset name.

model = VLAModel.from_preset("openvla-7b")

Parameters:

name (str) — Preset name: openvla-7b, smolvla-450m, dream-vla-7b

Returns: VLAModel instance

`VLAModel.from_checkpoint(path)`

Load a model from a local checkpoint.

model = VLAModel.from_checkpoint("./checkpoints/my-model")

`model.predict(image, instruction)`

Run inference on a single image-instruction pair.

Parameters:

image (PIL.Image) — Camera image
instruction (str) — Natural language instruction

Returns: Action object with fields: x, y, z, roll, pitch, yaw, gripper

Action

@dataclass
class Action:
    x: float       # Position delta x
    y: float       # Position delta y
    z: float       # Position delta z
    roll: float    # Orientation delta roll
    pitch: float   # Orientation delta pitch
    yaw: float     # Orientation delta yaw
    gripper: float # Gripper state (0=closed, 1=open)

VLATrainer

Training interface for fine-tuning VLA models.

`VLATrainer(config)`

Parameters:

config (TrainingConfig) — Training configuration

`TrainingConfig`

Field	Type	Default	Description
model	str	required	Preset name or checkpoint path
dataset	str	required	Path to HDF5 dataset
method	str	"lora"	Training method: lora, qlora, full
num_epochs	int	10	Number of training epochs
learning_rate	float	2e-5	Learning rate
batch_size	int	16	Batch size
lora_rank	int	16	LoRA rank
lora_alpha	int	32	LoRA scaling factor
lora_dropout	float	0.05	LoRA dropout
wandb_project	str	null	W&B project name
mixed_precision	str	"bf16"	Mixed precision mode

RobotController

Deploy a model to a real or simulated robot.

`RobotController(model, robot, control_hz)`

Parameters:

model (VLAModel) — Trained model
robot (str) — Robot name: widowx, franka_panda, ur5
control_hz (float) — Control loop frequency

Methods

start() — Initialize robot connection
execute(instruction, max_steps=50) — Run task
stop() — Clean shutdown

CLI

vlarobot presets                          # List available presets
vlarobot train --model <name> --dataset <path> --method <method>
vlarobot evaluate --model <name> --benchmark <name>
vlarobot predict --model <name> --image <path> --instruction "<text>"
vlarobot --version                        # Show version