| # exp_all_changes — what changed vs. main-branch FSM |
|
|
| This version applies **all four proposed fixes simultaneously** to compress the FSM |
| into the 20-second leaderboard eval window AND place objects more gently. |
|
|
| | Constant | Original (main) | This version | Why | |
| |---|---|---|---| |
| | `_PHASE_DURATIONS_PER_OBJECT` | `(180, 130, 20, 160, 170, 15, 30)` = 705 steps | `(80, 70, 25, 70, 100, 30, 25)` = 400 steps | Cuts wasted idle time. Original took ~25s per pair; new ~13s. Lets a learned policy finish inside the 20s eval cap. | |
| | `_LIFT_Z_OFFSET` | `0.20` | `0.15` | Reduces vertical travel during transit. (Earlier 0.10 attempt was too low — held cutlery clipped the plate edge.) | |
| | `_RELEASE_Z_OFFSET` | `0.09` | `0.06` | Cutlery dropped from 9 cm bounced out of the 15 cm success radius. Lower drop = gentler placement. (Earlier 0.04 attempt put gripper fingers into the table.) | |
| | `_MAX_CARTESIAN_DELTA` | `0.018` | `0.020` | Slightly faster IK so the cut phase durations are still enough for the actual motion. (Earlier 0.025 was too aggressive — caused knock-offs.) | |
|
|
| ## What the videos should show |
| - Faster overall execution (each episode ~13s vs ~25s) |
| - Lower transport altitude |
| - Cutlery placed more gently next to plate |
|
|
| ## What this experiment is **NOT** testing |
| - Original FSM behavior (see `exp_heights_only` for that baseline) |
| - Eval policy timing — these are FSM rollouts, not trained-policy rollouts |
|
|
