Physical AI platform
PLEXA One connects task intelligence, programs, perception, simulation, and learned controllers to a governed robot stack. Teams can extend the stack through the Control API, the Real-Time Control Interface, and the integration catalog.
For the software stack as a whole, explore Discover PLEXA OS · Coming soon and its three overlapping layers.

Two paths into robot execution
Control API requests pass through Robot Runtime for validation, planning, and execution. Direct RT clients send buffered paths through rt-msg. Give direct RT clients an explicit execution owner and coordinate them with Runtime commands before combining the two paths. Embedded robot controls retain functional-safety responsibility.
Three ways to drive a task
Compare scripted logic, bounded agent tools and learned policies using the same input. These development paths can be combined: an agent can start a program, and a program can use a policy. RT streaming is a separate execution path and requires an assigned execution adapter.
- 01 · INPUTTask + observation
- 02 · DECISIONWritten branches and parameters
if inspection == good: prepare_task() - 03 · EXECUTION BOUNDARYSDK → Runtime
The same task and observation enter each development path. Follow the next step.
The policy path shown here submits proposals through an application adapter. Direct RT streaming needs its own execution coordination; learned output alone defines neither that adapter nor the robot behaviour.
Integration depth
| Level | Use |
|---|---|
| 1 · Core application | Use the supported control stack for the agreed robot configuration and task. |
| 2 · Extend | Add AI, vision, data, or simulation around Runtime APIs and bounded capabilities. |
| 3 · Deep integration | Connect external robotics and policy stacks through reviewed adapters, Runtime admission, and embodiment-specific validation. |
Choose an adapter from the integration catalog and follow its setup, supported inputs and application examples.
Explore application tutorials
The tutorials organize common platform workflows by application. Start with teleoperation for continuous joint targets, ROS 2 for a middleware boundary, or Isaac simulation for a simulation workflow. Each outline points to the relevant API and contract; consult the integration catalog for implementation status.
Where learning helps
Deterministic programs and model-based motion provide repeatable operations. Teleoperation and motion tracking support teaching and demonstration capture. Simulation, imitation learning, and reinforcement learning support bounded problems such as grasping and adaptation to variation. Policy outputs require offline and simulator evaluation before supervised physical use.