Quad Training¶

Overview¶

This package provides the data-collection infrastructure for training learned policies on Quad-SDK. It handles time-synchronization of asynchronous simulation topics, episode lifecycle management, and a ready-gate that ensures the robot is fully initialized before data recording begins. Together these produce clean (state, action) rollouts suitable for offline RL, behavioral cloning, or dynamics model learning.

License¶

The source code is released under a MIT License.

Affiliation: Robomechanics Lab, Carnegie Mellon University

Tested under ROS2 Jazzy on Ubuntu 24.04.

Build¶

This is an ament_python package.

colcon build --packages-select quad_training

Usage¶

End-to-end data collection loop (spawns robot, waits for ready, issues commands, records synced bag, exits on termination):

ros2 launch quad_training data_collection.py

Lightweight variant that only launches the time-sync + bag recorder against an already-running stack:

ros2 launch quad_training training_logging.py

Nodes¶

time_sync_node¶

Subscribes to key simulation topics at their native rates and republishes aligned tuples on /synced/<topic> using message_filters.ApproximateTimeSynchronizer. The synchronized bundles are bagged as coherent (state, action) pairs for offline training.

Synchronized topics:

state/ground_truth (quad_msgs/RobotState, ~500 Hz)
control/joint_command (quad_msgs/LegCommandArray, ~500 Hz)
control/grfs (quad_msgs/GRFArray, ~500 Hz)
local_plan (quad_msgs/RobotPlan, ~333 Hz)
state/estimate (quad_msgs/RobotState, ~500 Hz)

Bundle rate is bounded by the slowest topic and the configured slop tolerance.

episode_monitor_node¶

Watches for three termination conditions and cleanly exits the launch context when any fire:

Goal reached — boolean on goal_reached.
Planner failure — boolean on planner_failed.
Body collision — roll/pitch exceeds a configured threshold.

After a condition triggers, the node waits a brief settle period to flush the bag, then raises SystemExit so the parent ros2 launch shuts down the full stack.

wait_for_robot_node¶

Blocks until the simulated robot is fully ready, so downstream data collection doesn't capture spawn-in artifacts. Exits only when all of the following hold for a configured consecutive-message count:

state/ground_truth is being published (Gazebo + driver up).
state/joints is being published (controllers activated).
Robot is upright (roll/pitch within threshold).
Robot has landed (vertical velocity near zero).

Launch Files¶

launch/data_collection.py — full rollout: bring up sim, wait for ready, start sync + bag, monitor episode, shutdown on termination.
launch/training_logging.py — sync + bag only, for attaching to an existing stack.

Bugs & Feature Requests¶

Please report bugs and request features using the Issue Tracker.