maum-ai/CostNav-Teleop-Dataset, measured from the public release on
16 September 2026.
Every row is read out of the release. Nothing here is estimated and nothing is rounded up.
| recordings | 1,639 |
| manifests read | 1,639 |
| topics declared in every recording and always empty | /robot_pose, /waypoints, /clicked_point, /odom/latency, /map_server/transition_event, /events/write_split |
| recordings missing /front_2d_lidar/scan and /back_2d_lidar/scan | 1,352 of 1,639, 82.5% |
| recordings missing the food camera | 822 of 1,639 |
| total duration | 40.73 hours |
| median recording | 81 seconds |
| messages across the corpus | 58,617,762 |
Measured from the recorder's own manifests. The mcap bags are 3.0 TB and were not downloaded, so this is a statement about what the manifests declare, not about decoded messages.
/robot_pose and /waypoints being empty everywhere means the navigation target and the robot's own pose estimate are not in the bags, so the corpus records what was driven without recording what it was driven towards. And a sensor suite that changes under you in four fifths of the corpus is not one input distribution.
Every bag ships a rosbag2 metadata.yaml beside it. Read topics_with_message_count and look at message_count per topic. No download of the bags needed and no credentials.
This page is what the release tells you on its own. A full grading removes the part that scores worst, retrains on what is left, and reports whether the smaller corpus scores higher, with rollout counts and intervals. The one result measured on physical hardware is Stera-10M, two arms on a four arm SO-ARM101 bench, 21.9% against 36.9% task success across 1,270 rollouts. The Stera-10M report.