roboloom
roboloom is a Python package that collects observations and actions by teleoperating a robot, saves them as a LeRobotDataset (0.7.0), and drives the same robot with the policy you train on them.
- Collect: drive the follower with a leader or a hand-held device and record it with cameras and tactile sensors.
- Train: train on the recording as it is, with
lerobot-trainor your own model. - Run and evaluate: let a policy drive the follower, take over when needed, and keep the verdict.
Each node runs as its own process in a dora-rs dataflow. A Session (the operator screen, the CLI, Python) starts the dataflow and sends episode start, stop and save.
flowchart LR
S["Session<br/>screen, CLI, Python"]
I["input<br/>leader, Pika Sense"] --> C["controller"]
P["policy"] --> C
C --> F["follower<br/>Rakuda2, PiPER"]
X["sensors<br/>RealSense, DIGIT"] --> R["recorder"]
F --> R
I --> R
R --> D[("LeRobotDataset")]
S -. "start, stop, save" .-> C
S -.-> R
-
Installation, concepts, the support matrix.
-
From a first run without devices to collecting, training and running a policy on Rakuda2.
-
Each robot's joints and the controllers and sensors it can use.
-
Writing a policy, LeRobot policies, interventions, evaluation.
-
Configuration keys, the CLI, the Python API, dataset columns.
Status labels
Each feature carries a label for how far it has been checked: Hardware-verified Mock only In development Design only. They are defined in Status labels. Use the hardware commands only after the operator has checked the values and the behaviour on each robot. A successful mock run or docs build is not a hardware check.