task01_picking_up_trash/code/PRIOR_DESIGN.md

Fixed-scene localization and navigation

The user explicitly authorized storing a map of the scene shared by all Task 01 instances. This policy uses that known-scene condition. It does not assume that the cans or bin remain in the same positions.

Offline preparation

debug/build_scene_prior.py fuses recorded RGB-D views using offline debug robot poses as survey calibration. Debug task-object poses only mask cans and the bin out of the survey. No task-object positions are exported. Static occupancy uses surfaces above 0.38 m; localization references use textured floor and furniture. The frozen atlas contains 57 views from 279 stable survey frames. Provenance is in assets/scene_prior_v1/survey_manifest.json.

Runtime

1. Extract ALIKED features from the official head RGB image and attach official depth/calibration to each feature.
2. Retrieve candidate reference views with a local visual-word histogram.
3. Match ALIKED features with LightGlue and fit a metric rigid SE(2) transform using RANSAC, vertical consistency, spatial support, and residual checks.
4. Initialize only after strong geometric verification. Between absolute fixes, use short RGB-D registrations and bounded proprioceptive velocity prediction.
5. Route on the frozen occupancy map plus the current depth obstacle layer. The live layer is rebuilt when observed, so localization corrections do not leave duplicate walls in the static map.
6. If tracking becomes stale, stop translation and search nearby viewing angles before retrying a route.
7. Prefer confirmed can proposals on unobstructed floor over furniture-overlap proposals, but retain the latter as fallback candidates for real cans under tables. Grasping and bin placement still refine geometry from live RGB-D.
8. When furniture blocks the nominal close grasp stance, check hover, descent, and lift IK from a farther collision-clear stance. The map clearance is not reduced to force an approach.

No evaluation-time robot/world pose, object pose, instance ID, score, simulator state file, backend request, or cloud model is passed to the policy. The frozen prior and pretrained weights are explicit offline inputs. This is known-scene evaluation, not a claim of unseen-scene generalization.

Algorithms and dependencies

• LightGlue, upstream commit eb42fee2d71449efb0aa5c10549752b5d75384d8, including ALIKED inference.
• ALIKED, aliked-n16 weights.
• Hierarchical Localization: reference retrieval followed by geometric localization informed the design; the hloc package itself is not a runtime dependency.
• RTAB-Map localization mode: fixed map with online pose correction informed the separation of mapping and localization; RTAB-Map itself is not a runtime dependency.

Vendored LightGlue retains its upstream licenses. Two loading branches were added to accept explicit local weight files; this adapter always uses those branches. Kornia 0.8.3 and kornia_rs 0.1.14 are installed under third_party/python, without changing the shared simulator environment. The latter wheel is for Linux x86-64 / Python 3.11. Torch and torchvision come from the existing environment. All weights are supplied under assets/models.

Validation

• Earlier replay after survey masking, before the repeated-floor rejection fix: 42/45 views localized; 37/45 met the stricter 40-inlier initialization threshold. Successful matches had median position error 3.23 mm, maximum 28.50 mm, and median localization time 34.4 ms. Network weight loading was disabled during this replay.
• Instance 301 navigation-only probe: 905 steps and 0.23 m base travel, final position error 3.35 mm, maximum 26.63 mm after localization. Estimates precede the action and debug poses follow it, so the live comparison includes that one-step offset.
• Instance 301 completed a fresh-reset development episode at Q=1.0 in 5,745 steps, with no task recovery. This used structural-support and prediction-consistency gates to reject repeated-floor aliases found in the preceding trial.
• Current-gate replay on the floor-alias failure sequence: 29/35 views localized, maximum position error 28.52 mm, with no accepted metre-scale aliases.
• Complete frozen standalone evaluation: all 20 official instances, mean final Q 0.766667, 13/20 official successes. See EVALUATION.md for individual scores and remaining failures. The diagnostic localization errors above are not a measured bound for the full suite.

Use --navigation prior in run.py for this policy. Frozen runs copy/hash the map, weights, dependencies, and source. The visual and validated modes remain available for comparison.

Placement now checks a copied joint-state sequence including preparatory hand rotation, the original bounded extension, vertical descent, and retreat before moving the base. If depth stops a partial correction, it rechecks the observed current stance. A trial with accumulated segmented torso extension regressed and was removed; absolute joint bounds and map/depth clearances remain unchanged.