A 5.31 × 10⁶ parameter monocular depth + 6-class segmentation student specialized for corridor-class indoor environments. The production corridor checkpoint of the Vortex-Depth lineage. Closed-loop validated against ground-truth depth in simulation.
LILocBench corridor RMSE: 0.382 m — the lineage's best corridor depth measurement.
Closed-loop navigation parity with ground-truth depth in Gazebo Fortress corridor experiments:
Depth source
Success rate (10 seeds)
Time-to-goal
Collisions
Ground-truth depth
9 / 10
17.77 ± 0.90 s
0
V9 (TensorRT FP16)
9 / 10
17.99 ± 0.45 s
0
The single failure is shared between configurations (identical seed, equivalent final pose error), indicating a planner-level failure mode rather than a depth-quality difference.
Use case
Recommended for fixed-route corridor deployment in indoor robot navigation. The intended deployment context is corridors with the structural characteristics of typical institutional indoor spaces: long parallel walls, mid-field-dominant depth distributions (1–15 m), uniform or fluorescent lighting. The model has been fine-tuned on the LILocBench Bonn corridor dataset and achieves competitive corridor depth metrics on that distribution.
For environments outside the corridor distribution, see:
At runtime, V9 depth is consumed by a confidence-gated fusion node combining ToF and student outputs per pixel:
s = median(d_tof / d_v9) over pixels where ToF confidence ≥ 0.5
if confidence(i, j) ≥ 0.5 AND 0.05 m < tof_depth(i, j) < 10.0 m:
fused(i, j) = tof_depth(i, j) # use hardware reading directly
else:
fused(i, j) = s · v9_depth(i, j) # use scaled learned depth
The per-frame median-scale step anchors the V9 prediction to metric units using the surviving ToF pixels. The model is therefore not required to produce absolute scale accuracy — it must produce accurate depth structure, which the runtime then anchors per frame.
Training
Fine-tuned from the V6 multi-domain pretrained checkpoint on LILocBench dynamics_0 for 50 epochs. AdamW optimizer, encoder LR 3 × 10⁻⁵, decoder LR 3 × 10⁻⁴, cosine annealing, batch size 16. Loss formulation unchanged from earlier lineage stages: berHu (depth) + cross-entropy (segmentation) + edge-aware smoothness, Kendall-weighted with log σ² clamped to [-2, 2].
The complete training pipeline that produces V9 spans four sequential stages:
Multi-domain pretraining on SUN RGB-D + DIODE Indoor + NYU Depth V2 (V6 pretrain checkpoint).
NYU Depth V2 fine-tuning with augmentation pipeline (V6 final checkpoint).
LILocBench dynamics_0 fine-tuning (V9 final checkpoint).
The first three stages are inherited from V6; only the final corridor-specialization stage is unique to V9.
Bootstrap perception context
This checkpoint is the production corridor model in a three-checkpoint family released as part of the Vortex bootstrap-perception pipeline. The pipeline addresses the operational reality that Time-of-Flight depth sensors lose ~78 % of their pixels on reflective indoor surfaces (polished floors, glass walls); the student fills the dead pixels with consistent learned geometry, and runtime fusion combines surviving sensor pixels with the student output.
