Skip to main content

​
What is Manipulation?

Manipulation covers everything a robot does with its end-effector: picking, placing, pushing, assembling, handing over, and dexterous in-hand manipulation. On Cyberwave, you build manipulation systems by pairing a robot arm with a digital twin, controlling it through teleoperation or AI policies, and orchestrating multi-step tasks with workflows.
Most manipulation projects on Cyberwave follow the same loop: simulate the task → teleoperate to collect demonstrations → train a policy → deploy to hardware.

​
Hardware that Supports it

SO101 Robot Arms

Open-source 6-DOF arms for desk-based manipulation, teleoperation, and imitation learning.

Universal Robot UR7e

Industrial collaborative arm for production-grade pick and place and assembly tasks.

Boston Dynamics Spot Arm

Mobile manipulation: combine the Spot quadruped with its arm for inspection and retrieval.

Custom Arms

Bring any URDF-described arm and connect it through a custom driver.

​
How You Build It

​
1. Test in simulation first

Spin up a digital twin in the Environment Editor, drop in target objects, and verify your motion plans against the MuJoCo physics simulation before any hardware moves.

​
2. Drive the arm

ModeUse it forReference
Dashboard controllersQuick manual moves, demosLive Teleoperation
Leader-follower teleoperationCollecting demonstration dataLive Teleoperation
Python SDKScripted sequences, tests, CIPython SDK
ML policyAutonomous executionML Models
from cyberwave import Cyberwave

cw = Cyberwave()
arm = cw.twin("so101-main")

with cw.affect("live"):
    arm.joints.set("shoulder_pan", 10, degrees=True)
    arm.joints.set("shoulder_lift", -30, degrees=True)
    arm.joints.set("elbow_flex", 60, degrees=True)
    arm.joints.set("wrist_flex", 45, degrees=True)
    arm.joints.set("wrist_roll", 0, degrees=True)
    arm.joints.set("gripper", 0, degrees=True)

​
3. Train a policy from demonstrations

Record 10–15 episodes of teleoperated demonstrations, then train a Vision-Language-Action model in the cloud and deploy it back to the arm as a controller policy. See ML Models for the training pipeline.

​
4. Orchestrate with workflows

Chain perception, planning, and motion into repeatable sequences with Workflows. A pick-and-place workflow typically looks like camera frame → object detection model → IK solver → motion node → gripper node.

​
Where to Go Next

SO101 Quickstart

Stand up your first manipulation arm.

Train a VLA Model

End-to-end tutorial: collect data, train, deploy.

Workflows

Compose multi-step manipulation tasks visually.