| # Plan C Implementation: Better Intervention Sampling |
|
|
| ## Goal |
| Improve intervention quality for more informative counterfactuals. |
|
|
| ## Changes to Generation |
|
|
| ### 1. Near-Miss Focus (50% of interventions) |
| - Small perturbations around expert action |
| - Gaussian noise: σ = 0.1 for position, σ = 0.05 for rotation |
| - Test both directions (±δ) |
|
|
| ### 2. Decision Boundary Exploration (30%) |
| - Actions at success/failure boundary |
| - Perturb only critical dimensions |
| - Keep gripper state same as expert |
|
|
| ### 3. Systematic Coverage (20%) |
| - Grid around expert action |
| - Cover action space uniformly |
| - Ensure diverse failure modes |
|
|
| ### Old Distribution (K=16) |
| - Expert: 1 |
| - Random: ~8-10 |
| - Near-miss: ~3-5 |
| - Structured: ~2-3 |
|
|
| ### New Distribution (K=16) |
| - Expert: 1 |
| - Near-miss (small δ): 8 (50%) |
| - Decision boundary: 5 (30%) |
| - Systematic grid: 2 (20%) |
|
|
| ## Expected Impact |
| - More informative comparisons |
| - Better gradient signal for ranking |
| - Clearer success/failure patterns |
| - **+2-3% performance gain** |
|
|
| ## Implementation |
| File: `scripts/generate_maniskill_lattice.py` |
| Function: `sample_candidate_actions()` |
|
|
| Add `--candidate-mode enhanced` flag. |
|
|
