LexHo/tartanimu-a3v21

TartanIMU Challenge (IROS 2026) — team Lexxxxx — model a3v21_s42 (second leaderboard entry)

Single unified model, one shared set of weights, for all four platforms (car, dog, drone, human). This repository is our second considered submission; the primary one (a3v20_s42, official 0.25878) is at https://huggingface.co/LexHo/tartanimu-a3v20. Each repository holds exactly one model and the exact prediction file it produced. Frozen weights: weights/tartanimu_a3v21_s42.pt (sha256 7d90badeb2873b1d8f3e13ffca21a624cf2b2ae8a5f838473a7ac44964f9c07b). submission.csv is the exact prediction file behind the claimed Kaggle score (md5 a53e196f6b67e8ea52733f4386b5c403; official scoring service: TartanIMU Score 0.26306, macro AVE 0.22232, macro ATE20 0.63665).

Run

pip install -r requirements.txt
python predict.py --data /path/to/tartan-imu-challenge-iros2026 --out submission.csv

Input: index/test_windows.csv + test/*.npz (raw 6-axis IMU only). No internet, no ground truth, no platform label. Measured re-execution of this repository as downloaded (CPU only, isolated environment, process-level socket block, empty cache): 1 min 10 s wall, 1.4 GB RSS; on one GPU < 1 min, < 1 GB VRAM. Cross-machine numerical tolerance vs submission.csv (produced on an RTX 5090): componentwise mean 1.8e-5, p99 1.6e-4, max 1.6e-3 m/s. SHA256SUMS covers every file (sha256sum -c SHA256SUMS).

Method (short)

Same main network as the primary entry, without the learned-INS sub-modules and recursion head, with a recording-level FiLM conditioning: dense random-offset 1 s windows → 14 input channels (raw IMU + an 8-channel decomposition about a slow complementary-filter "up" direction computed from the same IMU) → dilated 1-D ResNet trunk (5 tokens per window) whose blocks are modulated (FiLM: per-channel scale and shift) by a 35-dimensional descriptor of the whole recording (per-axis mean / spread / step roughness of the raw IMU, rest-window statistics, a 10-bin |acc| spectrum, gyro-rate percentiles, log duration — IMU only, no labels) → bidirectional GRU over 40-window chunks → linear head → body-frame velocity per window. Huber loss (β 0.05). Trunk initialised from a masked-IMU self-supervised pre-training on the released train+val IMU (no labels). Training augmentations: time dilation k ∈ [0.7, 1.5] (drone; identity-source racing family capped at 1.2) and, for the racing family only, translation scaling s ∈ [1.0, 1.8] applied to velocity labels and the gravity-removed specific force ("S-fast"); yaw augmentation about the estimated up (direction-only) for the identity source. Final weights = last epoch of a 120-epoch train+val run, seed 42 (pre-designated; seed range reported in the technical report).

Compliance declaration

1. Single model, one shared set of weights? Yes. One checkpoint, one forward pass per chunk; no per-platform experts. The FiLM descriptor is a function of the recording's own IMU computed inside the same forward pass; it is not a platform label and nothing is routed.
2. Platform classification, routing or specialization at inference? No. An auxiliary platform head exists only as a training loss; it is not used at inference.
3. Ensembling, checkpoint averaging or test-time augmentation? No. Single checkpoint, deterministic inference, overlapping-chunk averaging of one model's own predictions only (stride K/2 within a trajectory).
4. Test-set leakage avoidance, incl. model / checkpoint selection? Test ground truth never used. The final checkpoint is the fixed last epoch (no selection on any test signal). The self-supervised trunk pre-training used only the released train+val IMU. No external data or external pretrained weights are used in the submitted system. Development-time experiments with an external dataset (NeuroBEM) were run and are disclosed in the report; none of their weights or data enter this model. A public-leaderboard probe with a physics post-processing rule was made during development and is NOT part of this submission.
5. Development-time vs submitted system? Identical inference code; the submitted checkpoint differs from earlier leaderboard entries by the training augmentation (S-fast) and the recording-level FiLM. Model/checkpoint selection used the released validation split and two stress folds of the training split.