# DTR-R2 current Updated: 2026-09-06 Status: `DTR_R2_DYNAMIC_RETAINED` Primary Development environment: **self-built UE5 StreetLabV4**, per the user's 2026-09-06 migration decision. Use `tools/run_obstacle_research.py`; CARLA is retained for historical evidence and explicitly needed supplementary checks. The default live bank now uses 10 Hz observations and measured conservative visual envelopes. See [laboratory completion checks](unreal/UE_LAB_ACCEPTANCE_20260906.md) and the preserved [initial migration record](unreal/UE_PRIMARY_LAB_20260906.md). ## Capability question Can future obstacle occupancy intersecting the wearer's route produce stable `ONSET / HOLD / ESCALATE / CLEAR` events while missing evidence stays `UNKNOWN`? The unresolved contribution question is what collision-state information X94 adds beyond a raw motion baseline and the same temporal event smoothing. ## Retained evidence and baseline - **Public/JRDB X21:** same-live-track transport of an already authorized X13 component; six consumed sequences reached `5/6` CONTACT, 11 false segments, 45.45% Event F1, `3.061 s` median lead, and `8/18` dropout recovery. This is same-source Development; genuinely source-disjoint confirmation is pending. - **CARLA X73:** the latest complete source-disjoint synthetic Development confirmation. On C35, parent-hull reconstruction improved X72 by `+6 TP / +0 FP / +2.24 pp` frame F1, reaching `132/18/40` TP/FP/FN and `88.00/76.74/81.99%` precision/recall/F1. The same-map, scripted-motion, detector and evaluator boundaries remain; successors lack its confirmation. - **CARLA X94:** consumed post-hoc Development reference, `1478/84/417` TP/FP/FN and `94.62/77.99/85.51%` frame precision/recall/F1 across eleven cohorts. Its one-observation full-dropout bridge added six TP and no FP over X93; same-parent evidence and an unchanged valid issued plan remain required. - **Credible simple comparator:** in the consumed C35 raw-input pilot, Kalman CV + route tube + `0.60 s` hold tied X94 at 66.67% Event F1 and four false segments, with two versus ten fragmentation runs. X94 retained better frame F1 (`84.62%` versus `81.52%`), median lead (`2.80` versus `2.65 s`) and CLEAR (`0.00` versus `0.20 s`). This tradeoff has no fresh eleven-arm adjudication. - **X95 remains a challenger:** `DTR_CARLA_X95_CONSUMED_CROSS_VALIDATION_GATE_NOT_MET`. Its consumed replay gained `3.74 pp` Event F1 and removed five false segments versus X94, but lost `7.02 pp` frame F1 and added 30 fragment gaps. Simple `0.60 s` hysteresis reached 89.07% frame F1 on that replay; complexity has not established a win. ## Current bottleneck Frozen avoidance-only R1 reached `30/30` instance/witness source-stratum passes and a complete FIT_ONLY RGB/depth join. FINAL_A's depth server then exited before its first depth frame with `Shader compilation failures are Fatal`; FINAL_B RGB/depth never started. Status is `NOT_EVALUABLE_SOURCE_CAPTURE_INTERRUPTED`: nine complete sensor shards remain, no detector, fit, prediction or final method score was opened, and task-owned processes, ports and leases were released. The eleven-arm adapters have focused implementation checks only. This source failure changes no algorithm inheritance. The completed two-scene 720p probe passed 100 synchronized RGB/depth pairs. Synchronization alone did not improve capture time (`57.23 -> 59.13 s`). With both arms synchronized, fast lossless PNG reduced `63.28 -> 22.00 s` (2.88x); 400 independently decoded images exactly matched their raw pixels, with 8.96% more encoded bytes. This was one ordered short comparison excluding warmup, not statistical throughput or long-run stability evidence. It preserves depth bytes but does not validate metric-depth decoding or establish a shader-crash fix. A subsequent Development composite reused nine intact shards, but FINAL_A depth again hit a shader fatal before any payload. A separate DX11 probe reached RPC but failed camera warmup. Both runs ended without detector, fit or scores; the client now detects server death promptly. Neither failed run is reopened here. A separately identified launch profile requesting synchronous PSO compilation passed three cold starts (600 independently checked images), then all three missing shards (3,276 images) and all native joins. The composite source is now admitted as reused Development. This is bounded completion, not a permanent shader-fix claim. See [startup and method diagnosis](CARLA_CAMERA_STARTUP_20260905.md). Method preparation generated 910 FIT_ONLY detector frames but failed S03's six-frame dropout recovery condition: frame 29 is nearly black, has zero candidates and no measured collision credential. No fitting or final scoring ran. A consumed FIT_ONLY proposal with earlier 2/3/6-frame windows passed on that episode only; it neither rescues the failed run nor confirms method gain. Registration remains blocked by the existing input-fingerprint mismatch. The older eleven-cohort raw comparison cannot be reconstructed fairly: only C35 retains the required dense model/evaluator payloads. Derived tracks or recaptured pixels do not restore the original raw-input comparison. ## Next decision 1. **UE Development:** use fixed RGB-D replay for perception changes and V4 closed loop for motion changes. The measured motion reference is DEPTH_ONLY; candidate DTR remains an optional challenger. The new 10 Hz Development bank completed 32 actual branches: both sets of 8 straight controls passed, candidate depth succeeded 5/8 and candidate DTR 6/8, with no success regression. The extra success was late crossing, with nine changed-action frames. This is rerendered Development, not exact-pixel pairing or promotion. 2. **Next useful research:** investigate the retained late-stop and obstructed- bypass failures, and whether the late-crossing gain repeats under the same declared envelope. Do not restore the old undersized proxy to improve scores. Preserve conservative contact, goal, time and action-support reporting. All 81 actual 10 Hz batch-prefix checks matched the incremental full state; laboratory completion receipts are linked above. This task does not start another algorithm run or enlarge the frozen comparison. 3. **CARLA preservation:** park the pending dropout-window continuation and eleven-arm execution during the UE migration. Do not restore missing CARLA assets or recapture by default. Keep complete source, frozen method snapshots, old failures and prior algorithm inheritance. UE results do not inherit CARLA confirmation authority. This switch starts no X97, learner or held-out run; JRDB source-disjoint confirmation remains a separate requirement. ## Boundaries - Frozen R1 is not retried or reclassified as fresh confirmation. Preserve its seals, nine shards and crash evidence; no protected fit/final access follows from capture-engineering success. - C35-C41 cannot be rerun as confirmation. C8-C11 admitted no evaluable X31 occlusion source; N4 v1 cannot resume or retry. It is consumed incomplete, not a three-town result. Historical outcomes and source gates remain unchanged. - Consumed diagnosis is Development, not fresh authority. Keep frozen thresholds, source gates, lifecycle, association, seeds and denominators with their original results; changes belong to a separately identified Development version. - `UNKNOWN` and `NOT_EVALUABLE` are not `CLEAR`, negative evidence or safety. Wearer-global route conflict owns event correctness; component identity is diagnostic. Public replay and CARLA do not establish Android readiness, natural-distribution performance, user benefit, deployment or safety. - Uncommitted candidates and outputs remain WIP. Existing structured inheritance roles and historical verdicts remain authoritative; this compaction changes none. ## Evidence links - [X21 result](X17_X21_TRACK_CARRIED_COMPONENT_ANCESTRY_2026-08-29.md), [X73 confirmation](carla/DTR_CARLA_C35_X73_FRESH_CONFIRMATION_20260901.md), [X94 result](carla/DTR_CARLA_X94_CONSUMED_ELEVEN_COHORT_DEVELOPMENT_20260901.md). - [C35 raw-input pilot](DTR_BASELINE_RECKONING_C35_RAW_PILOT_20260905.md), [X95 result and simple controls](carla/DTR_CARLA_X95_CONSUMED_CROSS_VALIDATION_20260901.md). - [Frozen comparison design](DTR_FINAL_RECKONING_ROSTER_R1_20260905.md), [latest source execution and crash](DTR_FINAL_SOURCE_EXECUTION_20260905.md); the later execution record owns the design's current execution status. - [Completed RGB/depth and lossless-PNG probe](CARLA_RGBD_THROUGHPUT_20260905.md). - [Failed Development composite and DX11 probe](CARLA_FAST_COMPOSITE_SOURCE_20260905.md). - [Detailed ledger and reproduction](README.md), [formal research governance](../../../docs/formal/RESEARCH_GOVERNANCE.md). The full superseded current is preserved exactly in Git at `daf5720064d98a93b75336469d18e9a2fe0023e5:research/active/dtr-r0/CURRENT.md`. Use that history anchor and the existing result files for the X24-X94 trajectory; this page owns the present decision, not a second history ledger.