Skip to content

Observe, act and run

Robot.observe, Robot.act and Robot.run. One control loop for simulated and real robots.

A Robot is one body in a world or on a bus. You read it with observe(), move it with act(), and run a policy on it with run().

d03_one_loopd03_one_loop

from strands_robots import Robot

arm = Robot("so101")
obs = arm.observe()                 # joints, pose, camera images
arm.act({"shoulder_pan": 0.3})      # one control step
arm.act({"gripper": 1.0}, seconds=1.0)  # move there over 1 s
res = arm.run(lambda obs: {"elbow_flex": 0.4}, hz=50, steps=60)
print(res.stopped_by)               # "steps"

Robot(name) joins the default world, or the world you pass as sim=. Use position, orientation and keyframe to place it.

act()

Pass a {joint: target} dict or a vector in action_keys order. Joints you leave out keep their last target, so act({}) holds the pose. Targets are clipped to the joint limits.

With seconds=, the robot moves smoothly from its current pose and returns the pose it reached.

Method Does
hold() Stand still at the current pose
set_pose(position, orientation) Teleport (simulation)
set_gains(kp, kd) Set the PD gains
close() Remove the robot from its world

On a Feetech arm each write is capped at max_speed (default 2 rad/s). Pass max_speed=None to turn the cap off. Each read asks the servos for position only, as LeRobot does, so velocities and efforts are NaN (not measured). Pass read_motion=True to read velocity and load too, at 4 more bytes per servo on every read.

run()

run() takes a Policy, any object with act(obs) and reset(), or a plain function. Stop it with one of:

Argument Stops after
steps= That many control steps
seconds= That much simulated time
stop_when= Your function returns true

Pass realtime=True to match the wall clock in simulation. Before the first step, run() names any camera the policy needs and the robot lacks.

task = arm.start(lambda obs: {"elbow_flex": 0.2}, seconds=30)  # in the background
res = arm.stop()                                               # returns the RunResult

To run several robots together, pass a dict: arm.sim.run({arm: p1, g1: p2}, steps=20).

d08_policy_hooksd08_policy_hooks

Recording and replay

run(record="data/wave") adds the run as one episode of a LeRobot v3 dataset. Policy("data/wave?episode=0") replays episode 0. See Recording a dataset.

d09_record_replayd09_record_replay

Edit page