Skip to content

AI agents

A Robot() is a Strands tool. Hand it to an Agent and it drives the robot for you - picking actions from natural language.

from strands import Agent
from strands_robots import Robot

robot = Robot("so100")
agent = Agent(tools=[robot])   # one tool, 60+ actions

agent("Add a red cube on the table and pick it up")
# → calls robot.add_object(...) then robot.run_policy(...)

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

Add more tools

from strands_robots.tools import gr00t_inference, pose_tool

agent = Agent(tools=[robot, gr00t_inference, pose_tool])
agent("Start a GR00T server on port 5555 with so100_dualcam, "
      "then pick up the cube using groot")

Everything in strands_robots/tools/ is a @tool the agent can call directly.

Multi-turn

agent("Set up a blue ball and a red cube in opposite corners")
agent("Add a side camera looking at the workspace")
agent("Run the mock policy for 5 seconds and report which objects are still on the table")

Sim to real

robot = Robot("so100", mode="real",                                    # requires hardware
              cameras={"wrist": {"type": "opencv",
                                 "index_or_path": "/dev/video0", "fps": 30}})
agent = Agent(tools=[robot])
agent("Pick up the cube")

The agent sees the same tool spec in both modes; only the implementation changes.

Common patterns

Instruction Action chain
"Reset the world" reset
"Add a 5cm red cube" add_object(shape='box', size=[0.025]*3, color=[1,0,0,1])
"Take a picture" render
"Run the policy" run_policy(robot_name='so100', ...)
"What's in the scene?" list_objects + get_state
"Try 10 episodes, report success" eval_policy(robot_name='so100', n_episodes=10)
"Record a session" start_recording → run_policy → stop_recording

Inspect the full spec: print(robot.tool_spec) - JSON schema with all 60+ actions.

See also