ai_final_7_eval_act_policy_1

ACT policy trained on the evaluation-aligned cup-stacking dataset.

Checkpoint revisions

Each checkpoint-N revision contains the complete pretrained_model saved after N optimizer training steps. A larger number means the model was trained for more update steps; it does not necessarily mean that its evaluation success rate is higher.

Revision	Training steps
checkpoint-100000	100,000
checkpoint-140000	140,000
checkpoint-180000	180,000
checkpoint-200000	200,000 (last)

Use the same fixed evaluation seed and number of rounds when comparing revisions.