| # DART-style perturbed demonstrations, merged after another profile: | |
| # PHASE2_PROFILE=baseline,dart | |
| # In a share of episodes the executed approach drifts off-centre by a random sideways offset | |
| # and returns to the plan during the descent, while the recorded actions stay the clean | |
| # plan (Expert.perturb_pick). The policy then sees off-centre states labelled with the | |
| # centred path, which the unperturbed teacher never shows: in its demonstrations the hand | |
| # always starts the descent exactly centred and goes straight down, so the action near the | |
| # piece can be predicted from the joint state alone (sim/reports/replan_test). | |
| [dataset] | |
| repo_id = "local/so101_chess_baseline_dart" | |
| root = "data/so101_chess_baseline_dart" | |
| [expert] | |
| dart_probability = 0.5 # share of episodes perturbed | |
| dart_radius_mm = [2.0, 10.0] # offset size, uniform, any direction (the baseline policy's | |
| # error 40 mm above the piece: ~4 mm in successes, ~8 mm in failures) | |
| dart_clear_mm = 4.0 # back on the plan this far above the target's top | |
| dart_return_s = [0.3, 0.8] # the return takes this long (uniform), so off-centre states reach different heights | |
| dart_tries = 4 # offsets drawn before running the episode unperturbed | |