SO101 centered rack approach
Exploratory ACT policy trained locally on two camera views and six recorded joint positions.
Dataset: Histochemichael/so101-rack-centered-approach-2cam-25. 20 training episodes; 5 held-out validation episodes. No independent test set. Target: open the gripper, approach the rack handle and hold alignment from the centered retracted start. This policy does not learn lifting or insertion.
Current optimizer step: 25000/25000. Best validated checkpoint: 24000. The root model files, once present, contain the best checkpoint by held-out inference L1, not necessarily the final step. Policy cadence: 15 Hz nominal, 30-action prediction chunk, 3 actions per policy call. Original data were encoded at 30 FPS despite 15 Hz camera collection; a separate losslessly retimed copy is used. Exact physical capture timing is not recoverable. Do not assume nominal timing proves synchronization.
State/action normalization uses training episodes only. Validation runs in inference mode without target actions and averages errors over valid, non-padding actions across every held-out frame. Training total loss is a logged interval mean; training L1 and KL entries are latest-batch values. Offline errors are not physical grasp success rates.
Dashboard snapshots are checked every minute. Reload this page for the latest published snapshot.