Replay: python state-estimation/replay/replay_field_ekf.py → a Python port of the C++ EKF tracks the recorded field pose to a median of 0.008 mm after the cold start (130 tag updates; 1 of 130 above the 95 % line of NIS, the normalized innovation squared).
The AprilTag localizer turns a camera image into the camera’s pose on the field, plus a 6x6 covariance.
With two or more tags in view, all corners are pooled. If the geometry is not observable, the pose is withheld instead of inventing a covariance.
The point cloud is the camera driver’s own output; the localizer started from a teammate’s port, and its pooled multi-tag solve and covariance are mine.
Replay: python perception-sqpnp/replay/replay_localizer.py → re-solving 178 pooled frames from the logged corners, which the node rounded to whole pixels, lands within a median 14.1 mm of the recorded pose; from unrounded corners reprojected from each tag’s logged pose, 0.48 mm.
Foxglove buttons open the clip in a browser viewer.
Shoot on the move
Clip: the competition robot shooting while it drives.Open the MP4 file
A turret that keeps aiming while the robot drives, with shot compensation for the robot’s velocity.
A ball fired from a moving robot keeps the robot’s velocity. The compensation looks ahead and splits the velocity into the part toward the target and the part across it.
I wrote the turret’s aiming cascade and its shot compensation, with contributions from teammates.
Replay: python control/turret-2026/replay/replay_shot_solution.py → the turret loop’s aim error and feedforward recompute from the 30 s log to within 0.0010° and 0.0005 V; tracking RMS 0.90° (in this clip the turret held a fixed field heading and shot compensation was off; the compensation law is shown as a counterfactual).
A solver on the coprocessor decides how the whole robot moves.
A phase state machine runs DRIVE, APPROACH, TASK and RETRACT. On entering TASK, the goal is converted into the solver’s frame and frozen, so it stays at one field point.
The goal freeze and the QP moving the chassis were simulation-only.
Reaching a field-fixed goal from AprilTag localization
Clip: the arm following an AprilTag held at the end of a stick.Open the MP4 file
AprilTag fixes tell the robot where it stands, so the arm’s goal can be fixed on the field.
The arm’s goals are poses fixed on the field, while the whole-body solver works in the robot’s own frame. Before TASK, APPROACH requires the field pose to be fresh.
This demo mode is not excerpted in the showcase repository.