SO-101 single arm build - how to train ACT without a second leader arm + best CPU simulation alternatives to Isaac Lab?
​
Hey everyone,
Building a SO-101 6-DOF arm for autonomous pick and place with drop recovery. Using LeRobot + ACT policy + ROS2 Jazzy on Ubuntu 24.04.
my setup -
- Single SO-101 follower arm (can't afford the leader arm)
- Lenovo i3 laptop, Intel UHD only, no NVIDIA GPU
- PyBullet and MuJoCo working, Isaac Lab is out for me
wanted to know -
1.Single arm training : LeRobot normally needs leader + follower. Has anyone trained ACT with just one arm? Keyboard teleoperation? Gamepad? Sim-to-real from MuJoCo?
Simulation without GPU— Isaac Lab is unusable on my machine. Is Webots or Genesis viable on Intel UHD? Any ROS2-friendly sim that actually works on CPU?
Virtual demo collection — any tool or GitHub repo that lets you move a virtual arm with keyboard/mouse and export as LeRobot-compatible dataset?
Drop recovery— using STS3215 servo load register + YOLOv8 wrist camera fusion to detect drops, then FoundationPose for re-grasp. Has anyone done anything similar on cheap hardware? Any gotchas?
Any GitHub repos, Discord servers, or tips appreciated 🙏
Stack: ROS2 Jazzy | LeRobot | ACT | PyBullet | MuJoCo | YOLOv8 | FoundationPose | MoveIt2