robots

One call, two targets

Strands Robots is built on top of the Strands Agents harness and gives an agent physical capabilities through a single Robot constructor. Calling Robot with a model name such as so100 returns a MuJoCo simulation by default, which needs no GPU and no hardware, while passing mode equals real targets an actual robot with the same code and the same natural language control. An optional peer to peer mesh can be enabled so that multiple robots can coordinate. The design goal is that the mental model stays identical whether you are working in a simulation or on a bench with a real arm.

The robotics loop in practice

The library covers the whole robotics workflow in one place. You can teleoperate a real leader arm and record demonstrations into a LeRobotDataset, fine tune a policy on those demos with the train_policy helper using a LoRA method and a local LeRobot backend, then run the resulting checkpoint both in the MuJoCo twin and on the physical arm. A fleet peer can be told to assist in natural language through the mesh interface. Cameras are attached per robot, and the same policy file is reused across simulation and hardware so that behavior stays consistent.

Ecosystem and licensing

The project integrates with several robotics and AI ecosystems: MuJoCo version 3.x for simulation, NVIDIA GR00T, Hugging Face LeRobot for datasets and policies, and ROS 2 for broader robot middleware. Documentation and a project board are linked from the README, and a separate robots sim repository provides the simulation side. The code is released under the Apache-2.0 license and the August 2026 snapshot shows 128 stars, 31 forks, and 90 open issues, indicating active development.

Editorial conclusion

Strands Robots is Apache-2.0 licensed and reports 128 stars with 31 forks as of August 2026. It ties together MuJoCo simulation, LeRobot policy training, and real hardware control behind one agent friendly interface.