SO-101 robot arm: first autonomous pick-and-place
An ACT policy trained on my own teleoperation demos picks up a green cube and places it on a coaster. Dataset, model, and code stay private.
Key facts
- Pipeline: Hardware, teleoperation, data collection, training, and deployment, built and run by me.
- Policy: ACT, LeRobot 0.6.1, ResNet-18, chunk size 100.
- Data: 22 teleop demos, 9,400 frames.
- Training: 25k steps on a free Kaggle T4.
- Inputs: 6 joint positions plus overhead and wrist cameras, 480x640 at 15 fps.
- Inference: CPU on my Ubuntu machine, about 300 ms per action chunk.
- Result: With the final settings, 2 of 3 tries completed the pick-and-place (2 of 6 counting tuning runs).
- Limits: Placement is still imprecise, and success depends on where the cube starts. The dataset, model, and code are private.
- Date: Nov 2026.
I built and ran the whole pipeline myself: hardware, teleoperation, data collection, training and deployment. An ACT policy (LeRobot 0.6.1, ResNet-18, chunk size 100) was trained on 22 teleop demos (9,400 frames) for 25k steps on a free Kaggle T4. Its inputs are 6 joint positions plus overhead and wrist cameras (480x640 at 15 fps). Inference runs on CPU on my Ubuntu machine, at about 300 ms per action chunk. With no leader arm, the arm picks up a green cube and places it on a coaster. With the final settings, 2 of 3 tries completed the pick-and-place (2 of 6 counting tuning runs). Placement is still imprecise, and success depends on where the cube starts. The dataset, model and code are private. Nov 2026.