> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cyberwave.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Ways to control a robot
> Five ways to control your robots: manual input, skills, workflows, VLA models and RL policies. What decides each action, and when to use which.
Cyberwave gives you five ways to control the robots in an environment. They are alternatives, not steps: pick the one that fits the task, and switch whenever you like. Each drives the same twins, in simulation or on hardware, and exactly one **controller** drives a given twin at any moment. Switching between them doesn't stop telemetry or recording, so a session can start manual, hand over to a model, and fall back to a person if something goes wrong.
```mermaid theme={null}
flowchart LR
M[Manual input
keyboard · gamepad · leader arm] --> E
S[Skills
named actions] --> E
W[Workflows
automation] --> E
V[VLA models
learned from demonstrations] --> E
R[RL policies
learned in simulation] --> E
E[Robots in your environment
one controller per twin at a time]
```
| Way to control | Who decides the next action | Use it for | Start here |
| ---------------- | ----------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------- | -------------------------------------------------------------------- |
| **Manual input** | A person, with keyboard, gamepad or a leader arm | Demos, recovery, recording demonstrations | [Teleoperation](/feature-reference/environment-editor/teleoperation) |
| **Skills** | A named action the robot already knows: a saved pose, a standard command, a trained skill | "Go home", "open gripper", "pick", called by a person, a workflow or an agent | [Standard arm commands](/feature-reference/arm-commands) |
| **Workflows** | Rules you draw: trigger → model → decision → command | Repeatable automation, alerts, inspections | [Workflows](/feature-reference/workflows) |
| **VLA models** | A vision-language-action model, from camera frames and an instruction | Manipulation learned from demonstrations | [Fine-tune SmolVLA](/feature-reference/ml-models/smolvla-training) |
| **RL policies** | A policy trained by reward in simulation | Locomotion and dynamic control | [RL tasks](/feature-reference/rl-tasks) |
## Assign a controller
* **In the dashboard:** select the twin and use **Assign Controller**. A person can take over from the same panel at any time.
* **From Python:** `twin.policy.list()` shows the controllers available to a twin, `twin.policy.assign(policy)` hands it over, and `twin.policy.ensure_attached()` lets your script drive it.
* **From an agent:** the [control agent](/ai/control-agent) plans an action or a skill run, and you approve it before it executes.
How controllers are routed, where each type runs, and how code controllers and RL checkpoints run on a controller host: [Controllers and handover](/feature-reference/online-controllers).
## Where the models come from
Manual sessions, skills and workflows all produce recordings. Recordings become datasets, and datasets train VLA models and RL policies. See [Data for training](/feature-reference/datasets/import).
## Next steps
A camera, a model and an SO-101 twin that picks and sorts.
Record demonstrations, fine-tune, deploy the model as the controller.
Triggers, models and robot commands on a canvas.
Let an assistant plan the action, then approve it.