Skip to content

Strands Robots

Strands Robots: perceive, reason, act, world - the closed control loop around a Strands Agent core

A robot library for Strands agents. You name a robot, you get something you can drive - in simulation by default, or on real hardware when you ask for it.

from strands import Agent
from strands_robots import Robot

robot = Robot("so100")
agent = Agent(tools=[robot])
agent("Pick up the red cube")

Robot("so100") gives you a MuJoCo simulation - no GPU, runs on a laptop. Add mode="real" and the same code drives a physical arm. The agent sees one tool with 60+ actions; it figures out which to call.

graph LR
    A["pick up the cube"] --> B[Agent]
    B --> C["Robot(...)"]
    C -->|"mode=sim"| D["Simulation (MuJoCo)"]
    C -->|"mode=real"| E["HardwareRobot (LeRobot)"]
    D --> F[Policy]
    E --> F
    F --> D
    F --> E
    classDef agent fill:#0969da,stroke:#044289,color:#fff
    classDef hw fill:#bf8700,stroke:#875e00,color:#fff
    classDef policy fill:#8250df,stroke:#5a32a3,color:#fff
    class B,C agent
    class D,E hw
    class F policy

The agent decides what to do. The policy (Mock, GR00T, LeRobot, or Cosmos 3) decides how. The backend - physics or servos - does it.

Start here

  • New? Quickstart gets a robot moving in five minutes.
  • Building something? Quickstart gets a robot picking up a cube; the simulation and policy pages take it from there.
  • Looking for a robot? 68 of them, every one addressable by name.
  • Want the shape of it? Architecture is one diagram and a table.

Install

uv pip install "strands-robots[sim-mujoco]"   # simulation
uv pip install "strands-robots[all]"          # sim + hardware + every policy