fastlabel/open-lutra

OpenLUTRA is a data-recording system for robots that run ROS2, which is a widely used framework for building and controlling robots. The project is aimed at teams collecting training data for robot learning, a process sometimes called robot teaching. When teaching a robot to perform physical tasks, you typically need to record what the robot sees and does during demonstrations so that a model can learn from those recordings later. The system captures data streams coming from a robot, referred to as topics in ROS2, and saves them as MCAP files. MCAP is a file format designed specifically for storing timestamped robotics data. Rather than writing a custom recording engine, OpenLUTRA delegates the actual recording to the standard ROS2 bag tool and uses a separate lightweight process to monitor topic quality in real time. A web-based interface sits on top of all of this. From the browser, you can select which topics to record, start and stop recordings with a single button, and watch live statistics such as the data rate per topic and any gaps or missed messages. When a recording stops, the system automatically runs a quality analysis and flags topics where data loss exceeded a threshold. You can then browse past recordings, inspect per-topic timelines, and generate video previews from camera streams stored in the MCAP file. The project also provides a way to define custom validation rules, so teams can encode their own checks for what a valid recording looks like beyond the built-in metrics. Robot configurations, including which topics to monitor and at what expected rates, are defined in YAML files and switched via an environment variable. The project is at version 0.1.0 and is described as beta. APIs and data formats may change before a stable release. Docker images are not published yet, so the project is installed from source. External code contributions are not accepted at this stage, but bug reports and feature requests in English or Japanese are welcome through the GitHub issue tracker.