Skip to content

Quickstart

uv pip install "strands-robots[sim-mujoco]"

from strands_robots import Robot
import imageio.v3 as iio

sim = Robot("so100")                              # sim by default; CPU-only, no GPU
sim.step()

frame = sim.get_observation("so100")["default"]   # uint8 HxWx3
iio.imwrite("first_frame.png", frame)

Headless box? export MUJOCO_GL=osmesa before importing. See Troubleshooting.

Add an object and run a policy

sim.add_object(
    name="cube", shape="box", size=[0.025]*3,
    position=[0.3, 0.0, 0.025], color=[1, 0, 0, 1],
)

sim.run_policy(
    robot_name="so100",
    instruction="pick up the red cube",
    policy_provider="mock",   # no GPU; swap "groot" or "lerobot_local" with a real model
    duration=10.0,
)

Drive with an agent

from strands import Agent
from strands_robots import Robot

robot = Robot("so100")
agent = Agent(tools=[robot])
agent("Add a red cube and pick it up using the mock policy")

The whole loop

Teleoperate a real arm to collect demos, post-tune a policy on them, run it in sim and on hardware, hand work to a fleet peer, expose it on ROS 2 - one library. Each step is a distinct capability; the pages linked cover the details.

from strands import Agent
from strands_robots import Robot
from strands_robots import train_policy

# 1. TELEOPERATE a real SO-101 with its leader arm and RECORD demos.
follower = Robot("so101", mode="real", port="/dev/ttyACM0",
                 cameras={"front": {"type": "opencv", "index_or_path": "/dev/video0"}},
                 mesh=True)
follower.attach_teleop("so101_leader", port="/dev/ttyACM1", id="leader")
Agent(tools=[follower])(
    "start_recording(repo_id='me/pick', root='/tmp/pick', fps=30, "
    "task='pick up the cube'); teleoperate for 60s; stop_recording"
)

# 2. POST-TUNE a policy on those demos (LoRA fine-tune; GPU box).
train_policy(action="train", provider="lerobot_local",
             dataset_root="/tmp/pick", base_model="lerobot/smolvla_base",
             output_dir="/tmp/pick_ckpt", method="lora", steps=20000)

# 3. RUN the tuned checkpoint - same policy on a MuJoCo twin AND the real arm.
twin = Robot("so101")
twin.run_policy(robot_name="so101", policy_provider="lerobot_local",
                policy_config={"pretrained_name_or_path": "/tmp/pick_ckpt"}, duration=10.0)
follower.start_task("pick up the cube", policy_provider="lerobot_local",
                    policy_port=None, duration=10.0)

# 4. COORDINATE a fleet - tell a mesh peer to assist, in natural language.
follower.mesh.tell(follower.mesh.peers[0]["peer_id"], "hold the tray steady")

# 5. EXPOSE the running sim on ROS 2 - rviz / nav2 / any ros2 node can subscribe.
from strands_robots.simulation import Simulation
sim = Simulation(ros2_bridge=True); sim.create_world(); sim.add_robot("so101")
sim.step(100)
  1. Teleop + recording - Teleoperation, Recording.
  2. Post-tuning - Training.
  3. Sim and hardware rollout - Policies, Robot control.
  4. Fleet coordination - Mesh.
  5. ROS 2 interop - ROS 2.

Steps 1 and 3-real need hardware; step 2 needs a GPU. Everything else runs in sim.

Next: the notebook series

For a guided, click-and-run path, work through the getting-started notebooks, five notebooks that run end-to-end in simulation with no hardware, no GPU, and no Hugging Face credentials. They take you from Robot("so100") through recording a dataset, training a policy, and the full streaming data loop, each building on the last.

See also