Setup

Quick Start

Register your team. Join the competition on Kaggle and create your team there: TartanIMU Challenge: Multi-Platform Inertial Odometry. There’s no separate signup — your Kaggle team is your registration.

Download the train/val splits. Review the format and split details on the challenge Data page, then download the competition files from the Kaggle Data tab. Kaggle also provides index/test_windows.csv and sample_submission.csv.

Clone the baseline repository. superxslam/TartanIMU.

git clone https://github.com/superxslam/TartanIMU

Write a basic config file. For details on the available fields, see TartanIMU/config/datasets. This complete starter example keeps the optional training, evaluation, and fine-tuning settings visible:

Show complete starter config (minimal_example_config.yaml) 
# =============================================================================
# Minimal Example Configuration for Neural Inertial Tracking
# =============================================================================
#
# This configuration file provides a complete setup for:
# 1. Model inference on pre-trained neural inertial tracking models
# 2. Fine-tuning covariance heads for sensor fusion applications
# 3. Evaluation using the same metrics as the main test.py script
#
# Usage:
#   python minimal_example.py --config config/datasets/minimal_example/minimal_example_config.yaml
# =============================================================================

# =============================================================================
# MODEL ARCHITECTURE PARAMETERS
# =============================================================================
model_param:
  # Input/Output Configuration
  input_dim: 6                    # IMU channels: [acc_x, acc_y, acc_z, gyro_x, gyro_y, gyro_z]
  output_dim: 3                   # Body-frame velocity output: [vx, vy, vz]

  # LSTM Network Configuration
  lstm_size: 256                  # Hidden size of LSTM layers
  lstm_dropout: 0.0              # Dropout rate for LSTM layers (0.0 for inference)
  lstm_layers: 1                  # Number of LSTM layers
  num_layers: 2                   # Number of ResNet layers
  bidirectional: False            # Whether to use bidirectional LSTM

  # Output Head Configuration
  drop_ratio: 0.5                 # Dropout ratio for output heads
  split_z: True                   # Separate Z-axis prediction (recommended for 3D tracking)

  # ResNet Backbone Configuration
  layer_sizes: [2, 2, 2, 2]       # ResNet layer sizes for feature extraction

  # Time Window Configuration
  past_time: 0.0                  # Past time window (seconds)
  window_time: 1.0                # Current time window (seconds) - main prediction window
  future_time: 0.0                # Future time window (seconds)

  # Transformer Configuration (Advanced - usually not needed)
  use_transformer: False          # Whether to use transformer layers
  embed_dim: 256                  # Embedding dimension for transformer
  num_heads: 8                    # Number of attention heads
  num_blocks: 4                   # Number of transformer blocks

# =============================================================================
# TRAINING CONFIGURATION
# =============================================================================
train:
  # Sequence and Batch Configuration
  seq_len: 10                     # Sequence length for training (frames)
  batch_size: 32                  # Batch size for training
  debug: False
  learning_rate: 0.0001           # Learning rate (1e-4)
  epochs: 50                      # Number of training epochs

  # Hardware and Performance Configuration
  use_multi_gpu: False            # Whether to use multi-GPU training
  use_amp: True                   # Whether to use automatic mixed precision (recommended)

  # Training Strategy Configuration
  add_noise: True                 # Whether to add noise during training
  predict_start: 0                # Prediction start frame
  predict_end: 10                 # Prediction end frame

  # Covariance Training Configuration
  start_cov_epochs: 0             # Epoch to start covariance training (0 = from beginning)

  # Output Configuration
  out_dir: "/result/minimal_example"

  # Scheduler Configuration
  scheduler:
    factor: 0.5                   # Learning rate reduction factor
    patience: 10                  # Number of epochs without improvement before reducing LR
    min_lr: 1e-6                  # Minimum learning rate

# =============================================================================
# VALIDATION CONFIGURATION
# =============================================================================
val:
  batch_size: 32                  # Batch size for validation
  eval_freq: 5                    # Validation frequency (every N epochs)

# =============================================================================
# TEST CONFIGURATION
# =============================================================================
test:
  batch_size: 32                  # Batch size for testing
  out_dir: "/result/minimal_example_test"

  # Evaluation Configuration
  segment_length: 5.0             # Length of trajectory segments for evaluation (meters)
  compute_ate: True               # Whether to compute Absolute Trajectory Error
  compute_rte: True               # Whether to compute Relative Trajectory Error
  save_plots: True                # Whether to save evaluation plots

# =============================================================================
# LOGGING CONFIGURATION
# =============================================================================
logging:
  wandb_project: Neural_Inertial_Tracking_AirLab
  wandb_run_name: minimal_example

# =============================================================================
# DATA AUGMENTATION CONFIGURATION
# =============================================================================
augment:
  # IMU Noise Configuration
  feat_acc_sigma: 0.0001          # Accelerometer noise standard deviation
  feat_gyr_sigma: 0.00001         # Gyroscope noise standard deviation

  # Bias Configuration
  add_bias_noise: False           # Whether to add bias noise
  accel_bias_range: 0.1           # Accelerometer bias range
  gyro_bias_range: 0.002          # Gyroscope bias range

  # Gravity Configuration
  add_gravity_noise: False        # Whether to add gravity noise
  gravity_noise_theta_range: 5    # Gravity noise angle range (degrees)

# =============================================================================
# MODEL CONFIGURATION
# =============================================================================
model:
  model_name: "Foundation_Model"  # Model name for logging and identification
  model_yaml: "/neural_inertial_tracking/config/resnet_lstm_multihead.yaml"
  pred_velocity: True             # Whether to predict velocity (recommended)
  adapter: False                  # Whether to use adapter layers (for domain adaptation)

# =============================================================================
# DATA CONFIGURATION
# =============================================================================
data:
  # Dataset Configuration
  dataset: "AirLab"               # Dataset type (AirLab, KITTI, etc.)

  # Frequency Configuration
  imu_freq: 200.0                 # IMU frequency in Hz
  sample_freq: 40                 # Sampling frequency for training

  # Coordinate System Configuration
  use_local_coord: True           # Use local coordinate system (recommended for IMU)

  # Data Partitioning Configuration
  random_partition: False         # Whether to randomly partition data
  train_rate: 0.7                 # Training data ratio
  valid_rate: 0.15                # Validation data ratio
  test_rate: 0.15                 # Test data ratio
  data_rate: 1.0                  # Data usage ratio (1.0 = use all data)

# =============================================================================
# MOTION TYPE CONFIGURATION
# =============================================================================
motion_types:
  car: 1                          # Car motion type ID
  dog: 2                          # Dog motion type ID
  drone: 3                        # Drone motion type ID (most common)
  human: 4                        # Human motion type ID

# =============================================================================
# FINE-TUNING CONFIGURATION
# =============================================================================
finetune:
  # Covariance Head Fine-tuning
  covariance_lr: 1e-5             # Learning rate for covariance head fine-tuning
  covariance_epochs: 5            # Number of epochs for covariance fine-tuning

  # Debug Mode Configuration
  debug_mode: False               # Whether to enable debug mode (freeze velocity head)
  gradient_clip_norm: 1.0         # Gradient clipping norm for stability

  # Loss Configuration
  use_simple_loss: True           # Use simple MSE loss instead of complex loss
  covariance_target: 0.1          # Target covariance value for simple loss

Verify the model runs. Run an overfitting experiment on a small subset to confirm the training loop, data loading, and metrics are working end to end before scaling up.

Improve and submit. Experiment with your own methods to push performance on the leaderboard.

Log your submission and report. The Kaggle leaderboard alone doesn’t make you eligible for the final ranking or workshop presentation. Complete the Submission Form for each submission you want considered, and again to attach your technical report before the deadline. See Final ranking eligibility for the full requirements.

Schedule, Rules, and Leaderboard

Competition dates, team and submission limits, and any external-data policy can change during the event. Use the live Kaggle pages as the source of truth:

Timeline

  • Final Kaggle submission and model weights: September 20, 2026, 23:55 UTC
  • Technical report: September 23, 2026, 23:59 US Eastern Time (EDT, UTC−4) — three days after the leaderboard closes

Final ranking eligibility

The Kaggle leaderboard is provisional. Official final standings are decided only after the organizers re-execute the top teams’ submitted models. To be eligible for the final ranking you must, in addition to submitting predictions:

  1. Complete the Submission Form for each submission you want considered, attaching the prediction CSV you uploaded.
  2. Submit a technical report (PDF) — see Technical report below. All numbers in Tables III–V must come from the official scoring service, not a local re-implementation.
  3. Publish your single unified model’s weights and inference code as a public Hugging Face model repository, within the same 7 days.

Teams missing any of the three, or whose model cannot be re-executed, are removed from the final ranking regardless of leaderboard position.

Technical report

  • Template. Download the report template — a zip containing main.tex and ieeeconf.cls, plus a PDF preview. On Overleaf: New Project → Upload Project → select the zip, then recompile. The folder always holds the current version.
  • Format. Maximum 6 pages excluding references, and 7 pages in total including references, using the official template. The appendix does not count towards either limit.
  • How to submit. The report is not uploaded on this website. Fill in the challenge Form as usual and attach your report at the end of the Form — that single submission is what we collect.
  • Selection. The Kaggle leaderboard is not the only criterion for the Top 10: the quality of the report is also taken into account, and the Top 10 teams will be invited to contribute to the forthcoming IMU Foundation Model white paper.
  • Compliance declaration (required). Every report must explicitly answer: (1) do the predictions come from a single model with one shared set of weights? (2) is any platform classification, routing or specialization used at inference? (3) is any ensembling, checkpoint averaging or test-time augmentation used? (4) how was test-set leakage avoided, including for model and checkpoint selection? (5) what differs between the development-time and the submitted system? Reports that leave these unanswered cannot be considered for the Top 10.

Scoring your submission per sequence

The per-sequence scoring tool returns ATE20, AVE and RTE for all 89 test sequences, plus the per-platform and overall tables, using the same scoring logic as the Kaggle leaderboard. Five submissions per team per day; your team name must match your Kaggle team name exactly, and teams with no Kaggle submission are not scored.

RTE (Relative Trajectory Error, m) follows Sturm et al. (IROS 2012) and is the metric used in AirIO: over sliding windows of Δt = 5 s, it is the RMSE of the difference between the predicted and the ground-truth relative displacement. It measures local drift and, unlike ATE20, is insensitive to accumulated global offset. RTE does not enter the leaderboard ranking; it is reported for analysis only.

Submission format: one predicted body-frame velocity per test window, as a CSV with a header row:

window_id,vx,vy,vz
1,0.50,0.01,0.00
2,0.49,-0.00,-0.01

window_id values come from index/test_windows.csv; see sample_submission.csv on the Kaggle Data tab for the exact rows to fill in.

Reference Models

We provide per-platform “specialist” checkpoints trained internally as a benchmarking reference, so you can gauge what good performance looks like before submitting. You can also try them interactively in the live demo.

These specialists are a reference, not a submission strategy. Predictions must come from a single model with one shared set of weights. Inferring the embodiment inside your network is allowed and is precisely the point of the benchmark; recovering the platform in order to dispatch to separately trained per-platform experts is not. The test set is also anonymized with no platform label. Final rankings are subject to inspection of the submitted model and inference script. Use these checkpoints to sanity-check per-platform performance during development, not as your final submission.

  1. Start from a known baseline.
  2. Validate per platform first, then converge on a single unified model — final submissions must be one model scored across all four platforms with no platform label at test time.
  3. Profile both aggregate (macro-averaged ATE) and per-platform failure modes.
  4. Freeze your final model and produce reproducible inference logs.

System Requirements (Preliminary)

  • Linux or macOS development environment
  • Python + PyTorch toolchain
  • GPU required for both training and evaluation

Organizer Reference System

Results reported by the organizers were produced on the following hardware:

Component Specification
CPU AMD Ryzen 9 7950X (16-core / 32-thread)
Memory 128 GB DDR5
GPU NVIDIA RTX A6000 (48 GB VRAM)
Driver 535.183.01
CUDA 12.1
PyTorch 2.5.1+cu121