LightwheelAI/Lightwheel-Tasks-Double-Piper, 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.
| episodes declared in the release | 6,500 |
| episodes read at frame level | 1,687 |
| dimensions in observation.state | 16 |
| dimensions in action | 12 |
| arm joints named in observation.state | joint1 to joint6, both arms |
| arm joints named in action | joint1, joint2, joint3, joint5, joint6 |
| joint present in state and absent from action | joint4, on both arms |
| episodes in which joint4_l moves | 1,242 of 1,687, 73.6% |
| furthest joint4_l travels in one episode | 0.8715 radians, 49.9 degrees |
| episodes in which joint4_r moves | 620 of 1,687, 36.8% |
| furthest joint4_r travels in one episode | 0.1955 radians |
| median travel of the uncommandable joints | 0.0105 radians |
| median travel of the joints that are commanded | 0.5137 radians |
This is not a claim that the arm is still. Every other joint of the same naming series is commanded, so the gap is in the action space rather than in the behaviour. The check was also run against the so101 releases in the same account, where all six joints are commandable and there is nothing to report.
A policy trained on this action space cannot command joint4, while the observations it is trained against show joint4 moving in most episodes. Behaviour cloning then fits a five joint command to a six joint motion, and the residual has nowhere to go except into the other five joints.
Read meta/info.json and compare the names list under action with the names list under observation.state. The gap is visible without downloading a single parquet file.
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.