dtt-github · Nov 6, 2026 · Nov 6, 2026 · Nov 6, 2026 · Nov 6, 2026
diff --git a/README.md b/README.md
@@ -20,7 +20,7 @@ Initiated and maintained by **[DigitalTwin Technology GmbH](https://www.digitalt
 | Site | 128 m terrain with pad, haul road, stockpile and a 2.5 m excavation pit; 40+ physics props from NVIDIA SimReady (CC BY 4.0) with checked masses and colliders |
 | Real machines (open-licensed) | Hitachi ZX200 excavator, IHI (Kato) IC120 crawler carrier, Komatsu D37PXi dozer (articulated, from the PWRI OPERA project, Apache-2.0); 35 m tower crane, cones, barrels and jersey barriers (Gazebo, CC0) |
 | Soil | The pit filled with 2 m of MPM soil (Newton implicit MPM) on the GPU, two-way coupled to the machine, in four presets switchable live: dry sand, moist sand, loam and clay. Cohesion is checked against vertical-cut stability and McKyes cutting forces, and dug clods break up to their remoulded strength; only soil near the bucket is particles, resting soil elsewhere is a heightfield, and cut faces slump to the angle of repose. The sand steps at 30 Hz and the machine at 60 Hz; the dig cycle runs at 1.3× real time, and a trench dug in one lane (1.2×) reaches 1.7 m with 43–45° sides |
-| Excavator | The real Hitachi ZX200 (OPERA URDF geometry and joint limits, curved bucket with teeth) driven by its real hydraulic cylinders (Hitachi bores, 34.3 MPa main relief and port reliefs, pump flow shared between functions, bucket four-bar linkage) and swing motor (13.3 min⁻¹, 68.9 kN·m, braking on its cross-over reliefs), so force and speed vary with pose and the machine slows and stalls under load. Digging force at the teeth 149 kN against 151–158 kN published. Full dig–swing–load cycle into a dump truck, lifting while swinging like an operator: 15–18 s and about 1 t per pass |
+| Excavator | The real Hitachi ZX200 (OPERA URDF geometry and joint limits, curved bucket with teeth) driven by its real hydraulic cylinders (Hitachi bores, 34.3 MPa main relief and port reliefs, pump flow shared between functions, bucket four-bar linkage) and swing motor (braking on its cross-over reliefs, with friction), so force and speed vary with pose and the machine slows and stalls under load. Joint speeds per direction, swing speed (11.2 min⁻¹ at 1800 min⁻¹ eco) and swing friction are calibrated against PWRI's step tests on a real ZX200 (ISARC 2025); the swing torque (77.9 kN·m) and the bucket's angle offset are provisional ([phase2/RESULTS.md](phase2/RESULTS.md)). Digging force at the teeth 149 kN against 151–158 kN published. Full dig–swing–load cycle into a dump truck, lifting while swinging like an operator: 16–19 s per pass in a trench, about 20–21 s and 0.7–1.3 t per pass loading the carrier (clay lowest) |
 | Haul cycle | An IHI IC120 crawler carrier (PWRI OPERA model) is loaded by the excavator, drives the load to a spoil area, tips it out as MPM soil and comes back. The operator model sends it off when full, waits while it is away, and spreads passes along the vessel; sticky loads are shaken out. One excavator and one carrier deliver 124–181 t/h depending on soil, with the excavator waiting 16–26 % of the time. With two carriers (`--carriers 2`), they take turns at the excavator, a waiting place and gates keep them apart on the one-lane road, and loading goes up 12–27 % over 10 minutes. Over 25 minutes the gain holds at 19 % while the excavator has soil within reach ([phase2/RESULTS.md](phase2/RESULTS.md)) |
 | Spoil spreading | A Komatsu D37PXi dozer (PWRI OPERA model) spreads each dumped load in lanes fanned out from the dump heap and builds the fill layer by layer (0.3 m lifts), its blade pushing MPM soil under grade and load control as on the real machine; the tracks slip above traction. It works while the carrier is away and parks before it comes back, so the carrier never waits; mean blade forces of 5–28 kN per pass match the sliding-load estimate ([phase2/RESULTS.md](phase2/RESULTS.md)) |
 | Browser client | Levers and keyboard control, cylinder pressure gauges, stall warnings, live truck load, the carrier, the dozer and the spoil area, a "Send carrier" button; 60 fps |

diff --git a/phase2/RESULTS.md b/phase2/RESULTS.md
@@ -1483,32 +1483,47 @@ Haul logs: `logs/haul_bench_fill_long_<soil>_1carrier.json` (before) and `logs/h
 - **Clay passes come up light:** 0.81 t against the handbooks' 0.92–1.04 t and 0.93 t before. Loam moved from 5 % over its range to inside it (1.0–1.4 t).
 - **Clay and the face planner.** With the new hydraulics and the old planner, clay fell to 0.71 t a pass and one load took 25 passes. Per-pass traces showed no steady difference in what the bucket held at the start of the swing. The planner kept choosing one strip whose cut took slabs the bucket lost on the way out, at about 0.43 t a pass for some 30 passes. It brought back more than 30 % of a bucket at the start of the swing, so the planner never set it aside. Now, in clay, a pass that carries less than half a bucket to the carrier also sets its strip aside (`POOR_CARRIED`), and in every soil, strips set aside are taken again only when no other strip has a cut worth taking. That brought clay to 0.81 t and the longest load to 16 passes. The half-bucket rule hurt dry sand (1.11 t a pass): what spills on the swing there is no sign of a poor strip, so it applies to clay only. The remaining gap in clay is open.
 
-**Autopilot changes after the bucket calibration.** Two changes for clay, which loaded less with the new hydraulics:
-
-- **The curl gain isn't learned from a slab tear.** The operator learns how much the curl adds and curls that much short of a full bucket. In one 1500 s clay run, one curl lost 842 of 882 kg, and that single sample pulled the learned gain down. It ended at 18 kg, against 357 kg on the old hydraulics. So the bucket was curled only once overfull, and that tore out more slabs (11 against 6). A curl losing more than a quarter of the load (`CURL_TEAR`) is now counted (`curl_tears`) and not learned from. In that run the gain then stayed at 410–430 kg, and clay went from 0.81 to 0.85 t a pass.
-- **A poor strip is set aside at the carrier** (`_rest_if_poor`), not at the next approach. The next pass's strip is planned as the bucket swings back, so a poor strip used to be dug once more before its rest began. A log of the planner over 40 clay passes shows each poor strip set aside at once, and none taken again as a fallback.
+**The approach aimed at the previous strip.** Most thin clay passes came from the strip 20° left: the planner expected 0.8–1.2 m³ there, but its first pass brought out 3–59 kg, with three stages timed out. Neither the cut depth nor the drag was the cause. The pass count goes up only when the approach is reached, so the segment into the approach was planned with the previous pass's strip. The house swung back toward the new strip, turned to the old one for the approach, and then swung across to the new one again while penetrating. From −10° to +20° that is 37° of swing during the cut. The calibrated swing brakes at 0.82 rad/s², so it overshot to +27°, and the bucket stuck on the pit wall. A second pass in the same strip had no swing to make, which is why that strip dug well when chosen again. 'Swing back' and the approach now both aim for the next pass's strip (`_lane_pass`). The traced pass now carries 1.06 t with no stage skipped, and every pass that changes strip skips the detour.
 
 `bench_haul.py` now logs every pass (`pass_log`: strip, most held while digging, at the swing, carried, left after the dump) and stops after `PASSES=N` passes for a quick check.
 
-25-pass check against `main` (one run each, about 600 s; kg per pass, since t/h over 25 passes depends on where each run stops in a load):
+**Two other changes, tried and reverted.** Two other changes were tried for clay:
+
+- **The tear rule.** A curl that lost more than a quarter of the load would not be learned from (`CURL_TEAR`). One curl losing 842 of 882 kg had pulled the learned curl gain down to 18 kg.
+- **Setting a poor strip aside at the carrier** (`_rest_if_poor`) instead of at the next approach.
+
+Both were tested with the approach fix in place: 3 runs of 25 passes per variant. The figures are mean kg dumped per pass, with each run in brackets:
 
-| | Carried to the carrier | Dumped | Thin passes (under half a bucket at the swing) |
+| | Neither change | Both changes | Set-aside change only |
 |---|---|---|---|
-| Dry sand: main / with the changes | 1336 / 1332 kg | 1331 / 1329 kg | 1 / 1 |
-| Loam | 1220 / 1010 kg | 1116 / **906** kg | 3 / 6 |
-| Clay | 1180 / 1107 kg | 1025 / 973 kg | 2 / 3 |
+| Clay | **1041** (1071, 1085, 968) | 940 (1009, 844, 967) | 985 (1037, 969, 949) |
+| Loam | 1165 (1186, 1142, 1166) | 1093 (1103, 1112, 1063) | **1188** (1166, 1227, 1172) |
+| Dry sand | 1326 (1331, 1323, 1325) | 1331 (1338, 1313, 1342) | n/a: neither change acts in sand |
+
+- **The tear rule lowered both cohesive soils:** clay 5 % and loam 8 % below the set-aside change alone. So limiting it to clay would not help either.
+- **The learned gain no longer collapses:** it ended at 136–273 kg in the 9 clay runs.
+- **The set-aside change made no difference** in 25 passes, nor over 1500 s of clay. That is where the deep pit gives repeated thin passes from one strip: 55.5 and 53.0 t loaded with the change, against 52.3 and 58.4 t without it.
+
+Both are reverted.
+
+**The approach fix alone**, against the same autopilot without it:
 
-The check doesn't show the changes helping:
+- **25 passes:** clay and loam dump about 11 % more per pass (clay 935 → 1041 kg, mean of 4 old runs; loam 1047 → 1165 kg, 3 old runs), and passes are 1.5–2 s shorter.
+- **1500 s, clay (one carrier):** 43.5 → 52.3 and 58.4 t loaded, 0.65 → 0.73 and 0.85 t per pass (`loaded_t` over all passes), 104 → 125 and 140 t/h loaded. Stages skipped fell from 13 to 3 and 8.
+- **1500 s, dry sand:** no change beyond the spread: 70.0 and 62.2 t loaded before, 61.3 and 70.3 t after.
+- **1500 s, loam:** no change beyond the spread either: 69.2 and 70.0 t before, 63.8 and 69.4 t after. Every loam run dug out all the soil within reach before 1500 s.
+- **Other benches:** the trench (passes 1–6: 15.6–18.2 s) and the 130 s zone cycle (7 passes, 10.4 t) don't change strip between passes. `test_env.py` passes.
 
-- **Loam** dumped 19 % less with them. Two of its curls counted as tears, and its learned gain fell to −69 kg.
-- **Clay** dumped 5 % less, which is within the run-to-run spread.
-- **The collapse they guard against hadn't happened yet** on `main`: clay's learned gain was still 60 kg after 25 passes.
+**Still open:**
 
-Whether the tear rule suits loam, and whether either change helps over a long shift, needs several runs per soil.
+- **Clay still has thin passes once the pit is deep:** 17–26 of about 70 passes in the 1500 s runs. These passes barely fill while digging (160–600 kg at most), unlike the early failures, which filled and then lost the load. Half of them are in the strip 20° left.
+- **Dry sand stalls when a load is full.** When the last pass fills the carrier, the house sometimes stays over it, and the next 6 stages time out, losing about 30 s. This happened once to four times per 1500 s, with or without the fix, and sets most of dry sand's run-to-run spread.
 
-**Open: the clay strip 20° left.** The planner expects 0.8–1.2 m³ in that strip, but its passes bring out 3–59 kg. When its rest ends it is chosen again, and it fails again. Most thin clay passes come from it.
+Logs:
 
-Logs: `logs/haul_quick25_<main|fix>_<soil>_1carrier.json`. The 1500 s runs before the bucket calibration are `logs/haul_bench_pwri_face_long_<soil>_1carrier.json`, with the half-bucket rule in every soil, and `logs/haul_bench_pwri_tear_long_<loam|clay>_1carrier.json`, with the tear guard. `logs/haul_bench_pwri_rest_long_clay_1carrier.json` adds the earlier set-aside.
+- 25-pass runs: `logs/haul_quick25_lane_<base|fix|rest>_<soil>_<n>.json`; the earlier checks are `logs/haul_quick25_<main|fix>_<soil>_1carrier.json`.
+- 1500 s runs: `logs/haul_bench_lane_fix_long[2]_<soil>_1carrier.json`, `logs/haul_bench_lane_old_long2_<soil>_1carrier.json` and `logs/haul_bench_lane_restearly_long<n>_clay_1carrier.json`.
+- 1500 s runs before the bucket calibration: `logs/haul_bench_pwri_face_long_<soil>_1carrier.json`, with the half-bucket rule in every soil; `logs/haul_bench_pwri_tear_long_<loam|clay>_1carrier.json`, with the tear guard; `logs/haul_bench_pwri_rest_long_clay_1carrier.json`, which adds the earlier set-aside.
 
 **Limits:**
 - Every value is for 1800 min⁻¹ in eco mode. At full power the pumps deliver more: the spec swing speed is 19 % higher.

diff --git a/phase2/bench_haul.py b/phase2/bench_haul.py
@@ -117,7 +117,6 @@
     "pit_done": sim.pit_done,
     "faces": [(f["psi"], f["shift"]) for f in sim.face_log],
     "curl_gain_kg": round(sim.curl_gain),
-    "curl_tears": sim.curl_tears,
     "mean_carried_kg": round(float(np.mean([x["carried_kg"] for x in done]))) if done else None,
     "mean_dumped_kg": round(float(np.mean([x["carried_kg"] - x["left_kg"] for x in done]))) if done else None,
     "thin_passes": sum(1 for x in done if x["at_swing_kg"] < 0.5 * sim.full_kg),