AlgEngine Usage Guide

This guide covers how to use AlgEngine for training end-to-end autonomous driving models, running evaluations, extracting rare cases, and fine-tuning. AlgEngine is built on MMDetection3D and supports UniAD, VADv2, and HydraMDP architectures.

Quick Reference
Training
Evaluation
Rare Case Extraction
Fine-Tuning
Configuration
Model Architectures
Troubleshooting

Quick Reference

cd projects/AlgEngine
# Training (8 GPUs)
./scripts/e2e_dist_train.sh <config> <num_gpus> [resume_checkpoint]

# Open-loop evaluation
./scripts/e2e_dist_eval.sh <config> <checkpoint> <num_gpus>

# Rare case extraction
python scripts/rare_case_sampling_by_pdms.py \
    --pdm-result <csv_file> \
    --base-split <yaml_file> \
    --output-dir <output_dir>

# Closed-loop evaluation
bash scripts/run_ray_distributed_testing.sh <config> <checkpoint> <model_name> <data_type> <react_type>

Training

Prerequisites

Before training, ensure:

✅ AlgEngine environment installed (algengine conda env)
✅ Data prepared (see Data Organization)
✅ Pre-trained backbone weights downloaded

Training from Scratch

Train a model on 50% of the training data:

conda activate algengine
cd projects/AlgEngine

# Train VADv2 with 50% data (8 GPUs)
./scripts/e2e_dist_train.sh configs/worldengine/e2e_vadv2_50pct.py 8

Arguments:

<config>: Configuration file path
<num_gpus>: Number of GPUs to use
[resume_checkpoint] (optional): Checkpoint to resume from

Training with 100% Data

./scripts/e2e_dist_train.sh configs/worldengine/e2e_vadv2_100pct.py 8

Resume Training

Resume from a checkpoint:

./scripts/e2e_dist_train.sh \
    configs/worldengine/e2e_vadv2_50pct.py \
    8 \
    work_dirs/e2e_vadv2_50pct/latest.pth

Auto-resume: If latest.pth exists in work_dirs/, training will auto-resume.

Monitor Training

# Watch training log
tail -f work_dirs/e2e_vadv2_50pct/logs/train.*

# TensorBoard (if enabled)
tensorboard --logdir work_dirs/e2e_vadv2_50pct/tf_logs

Key metrics to monitor:

loss: Total training loss (should decrease)
loss_planning: Planning loss
loss_track: Tracking loss
ade_4s: Average displacement error at 4 seconds
fde_4s: Final displacement error at 4 seconds

Training Output

work_dirs/e2e_vadv2_50pct/
├── e2e_vadv2_50pct.py          # Config backup
├── logs/
│   └── train.26040614*         # Training logs
├── epoch_1.pth                 # Checkpoints
├── epoch_2.pth
...
├── epoch_20.pth
└── latest.pth                  # Symlink to latest checkpoint

Evaluation

Open-Loop Evaluation

Evaluate model predictions against ground truth trajectories.

Full Test Set Evaluation

conda activate algengine
cd projects/AlgEngine

# Evaluate on navtest (8 GPUs)
./scripts/e2e_dist_eval.sh \
    configs/worldengine/e2e_vadv2_50pct.py \
    work_dirs/e2e_vadv2_50pct/epoch_20.pth \
    8

Output:

work_dirs/e2e_vadv2_50pct/
└── navtest.csv                 # Evaluation results

Rare Navtest Cases Only

Evaluate on known rare navtest cases:

./scripts/e2e_dist_eval_navtest_failures.sh \
    configs/worldengine/e2e_vadv2_50pct.py \
    work_dirs/e2e_vadv2_50pct/epoch_20.pth \
    8

Output:

work_dirs/e2e_vadv2_50pct/
└── navtest_failures.csv        # rare navtest cases only

Understanding Evaluation Metrics

Open-loop metrics CSV format:

token,ade_4s,fde_4s,no_at_fault_collisions,drivable_area_compliance,ego_progress,comfort,score
abc123,0.42,0.85,1.0,0.95,0.88,0.92,0.89
...

Key metrics:

ade_4s: Average trajectory error over 4 seconds (meters, lower is better)
fde_4s: Final position error at 4 seconds (meters, lower is better)
no_at_fault_collisions: Collision avoidance rate (0-1, higher is better)
drivable_area_compliance: Stay in drivable area (0-1, higher is better)
ego_progress: Route completion (0-1, higher is better)
comfort: Comfort metric (0-1, higher is better)
score: Overall PDM score (0-1, higher is better)

Closed-Loop Evaluation

Evaluate model in simulation (requires SimEngine).

See SimEngine Usage Guide for:

Single-GPU testing
Multi-GPU distributed testing
Reactive vs non-reactive modes

Quick example:

cd projects/AlgEngine

bash scripts/run_ray_distributed_testing.sh \
    $WORLDENGINE_ROOT/projects/AlgEngine/configs/worldengine/e2e_vadv2_50pct.py \
    $WORLDENGINE_ROOT/projects/AlgEngine/work_dirs/e2e_vadv2_50pct/epoch_20.pth \
    e2e_vadv2_50pct_epoch20 \
    navtest_failures \
    NR

Rare Case Extraction

Extract failure scenarios from evaluation results for targeted fine-tuning.

Basic Extraction

conda activate algengine
cd projects/AlgEngine

python scripts/rare_case_sampling_by_pdms.py \
    --pdm-result work_dirs/e2e_vadv2_50pct/navtest.csv \
    --base-split configs/navsim_splits/navtest_split/navtest.yaml \
    --output-dir configs/navsim_splits/navtest_split/e2e_vadv2_50pct_rare

Arguments:

--pdm-result: CSV file with evaluation metrics
--base-split: Base scenario split YAML file
--output-dir: Directory to save extracted split files

Extracted Splits

The script generates three rare case split files:

configs/navsim_splits/navtest_split/e2e_vadv2_50pct_rare/
├── navtest_collision.yaml         # Collision scenarios
├── navtest_off_road.yaml          # Off-road scenarios
└── navtest_ep_1pct.yaml           # Low ego-progress (bottom 1%)

Using Custom Thresholds

Edit the script to customize:

# In rare_case_sampling_by_pdms.py

# Change collision threshold
collision_scenarios = df[df['no_at_fault_collisions'] < 0.95]  # From 1.0

# Change ego-progress percentile
ep_threshold = df['ego_progress'].quantile(0.05)  # From 0.01 (1% -> 5%)

Verify Extracted Scenarios

# Check how many scenarios were extracted
wc -l configs/navsim_splits/navtest_split/e2e_vadv2_50pct_rare/*.yaml

# View first few scenarios
head -20 configs/navsim_splits/navtest_split/e2e_vadv2_50pct_rare/navtest_collision.yaml

Fine-Tuning

Fine-tune a trained model on rare cases using reinforcement learning.

Prerequisites: Generating Rollouts with SimEngine

Important: Rollout data must be generated by SimEngine before fine-tuning. This involves:

Run SimEngine Rollout to generate trajectory data:

# Multi-GPU distributed rollout (recommended for large datasets)
export WORLDENGINE_ROOT=/path/to/WorldEngine
cd projects/SimEngine
bash scripts/run_ray_distributed_rollout.sh \
    $WORLDENGINE_ROOT/projects/AlgEngine/configs/worldengine/e2e_vadv2_50pct.py \
    $WORLDENGINE_ROOT/data/alg_engine/ckpts/e2e_vadv2_50pct_ep8.pth \
    e2e_vadv2_50pct \
    navtrain_50pct_collision \ 
    navtrain

Rollout Output is saved to:

experiments/closed_loop_exps/e2e_vadv2_50pct/navtrain_NR/
└── WE_output/
    └── openscene_format/
        ├── sensor_blobs/        # Camera images, LiDAR
        ├── meta_datas/          # Per-scenario metadata
        ├── pdms_pkl/            # Metric pdms pkl
        └── all_scenes_pdm_averages_NR.csv

Reorganize to AlgEngine Format (creates openscene-synthetic dataset):

conda activate simengine

cd projects/SimEngine
python scripts/export_simulation_data.py \
    --test_path experiments/closed_loop_exps/e2e_vadv2_50pct/navtrain_NR \
    --appendix 260406  # Date suffix for versioning, Default None

Output location: data/alg_engine/openscene-synthetic/

Verify Data Structure:

data/alg_engine/openscene-synthetic/
├── sensor_blobs/              # Replayed scenario sensor data
├── meta_datas/                # Metadata
└── pdms_pkl/                  # Metric pdms pkl

For detailed SimEngine usage, see SimEngine Usage Guide.

RL-Based Fine-Tuning

conda activate algengine
cd projects/AlgEngine

# Fine-tune on extracted rare cases (8 GPUs)
./scripts/e2e_dist_train.sh \
    configs/worldengine/e2e_vadv2_50pct_rlft_rare_log.py \
    8 \
    work_dirs/e2e_vadv2_50pct/epoch_20.pth

Arguments:

Config with _rlft_rare_log suffix (uses rare case splits)
Number of GPUs
Base checkpoint to fine-tune from

Fine-Tuning Configuration

The _rlft_rare_log config typically includes:

# configs/worldengine/e2e_vadv2_50pct_rlft_rare_log.py

# Use rare case splits
data = dict(
    train=dict(
        ann_file='merged_infos_navformer/nuplan_openscene_navtrain.pkl',
        scenario_filter=[
            'configs/navsim_splits/navtrain_split/e2e_vadv2_50pct_ep8/navtrain_50pct_collision.yaml',
            'configs/navsim_splits/navtrain_split/e2e_vadv2_50pct_ep8/navtrain_50pct_off_road.yaml',
            'configs/navsim_splits/navtrain_split/e2e_vadv2_50pct_ep8/navtrain_50pct_ep_1pct.yaml',
        ]
    )
)

# RL training settings
optimizer = dict(type='AdamW', lr=5e-5)  # Lower learning rate
total_epochs = 8  # Fewer epochs for fine-tuning

Fine-Tuning Output

work_dirs/e2e_vadv2_50pct_rlft_rare_log/
├── e2e_vadv2_50pct_rlft_rare_log.py
├── logs/
│   └── train.*
├── epoch_1.pth
...
└── epoch_8.pth

Evaluate Fine-Tuned Model

cd projects/AlgEngine
# Open-loop evaluation
./scripts/e2e_dist_eval.sh \
    configs/worldengine/e2e_vadv2_50pct_rlft_rare_log.py \
    work_dirs/e2e_vadv2_50pct_rlft_rare_log/epoch_8.pth \
    8

# Closed-loop evaluation
bash scripts/run_ray_distributed_testing.sh \
    $WORLDENGINE_ROOT/projects/AlgEngine/configs/worldengine/e2e_vadv2_50pct_rlft_rare_log.py \
    $WORLDENGINE_ROOT/projects/AlgEngine/work_dirs/e2e_vadv2_50pct_rlft_rare_log/epoch_8.pth \
    e2e_vadv2_50pct_rlft \
    navtest_failures \
    NR

Configuration

AlgEngine uses hierarchical configuration with MMDetection3D. For a detailed reference of all config parameters, variants, and their relationships, see the Configuration Guide.

Configuration Hierarchy

configs/
├── _base_/
│   └── default_runtime.py              # Base runtime settings
├── worldengine/
│   ├── e2e_vadv2_50pct.py              # 50% data training
│   ├── e2e_vadv2_100pct.py             # 100% data training
│   ├── e2e_vadv2_50pct_rlft_rare_log.py  # Rare case fine-tuning
│   └── ...
└── navsim_splits/
    ├── navtrain_split/
    │   ├── navtrain.yaml               # Full training set
    │   ├── navtrain_50pct.yaml         # 50% subset
    │   └── e2e_vadv2_50pct_ep8/        # Rare case splits
    │       ├── navtrain_50pct_collision.yaml
    │       ├── navtrain_50pct_off_road.yaml
    │       └── navtrain_50pct_ep_1pct.yaml
    └── navtest_split/
        ├── navtest.yaml                # Full test set
        └── navtest_failures.yaml       # Failure subset

Key Configuration Parameters

# Model architecture
model = dict(
    type='VADv2',  # or 'UniAD', 'HydraMDP'
    num_query=900,
    num_classes=7,
    planning_steps=8,
    img_backbone=dict(type='ResNet50', ...),
    img_neck=dict(type='FPN', ...),
)

# BEV configuration
bev_h_, bev_w_ = 200, 200
patch_size = [102.4, 102.4]  # Physical range (meters)

# Input modality
input_modality = dict(
    use_lidar=False,
    use_camera=True,  # 8 cameras
    use_radar=False,
    use_external=True  # CAN bus
)

# Training
total_epochs = 20
optimizer = dict(type='AdamW', lr=2e-4, weight_decay=0.01)
lr_config = dict(
    policy='CosineAnnealing',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=1.0 / 3,
)

# Data
data = dict(
    samples_per_gpu=1,
    workers_per_gpu=4,
    train=dict(
        ann_file='merged_infos_navformer/nuplan_openscene_navtrain.pkl',
        scenario_filter='configs/navsim_splits/navtrain_split/navtrain_50pct.yaml',
    ),
    val=dict(
        ann_file='merged_infos_navformer/nuplan_openscene_navtest.pkl',
        scenario_filter='configs/navsim_splits/navtest_split/navtest.yaml',
    ),
)

Override Configuration

Override config parameters at runtime:

./scripts/e2e_dist_train.sh \
    configs/worldengine/e2e_vadv2_50pct.py \
    8 \
    --cfg-options \
    optimizer.lr=1e-4 \
    total_epochs=30 \
    data.samples_per_gpu=2

Model Architectures

AlgEngine supports multiple end-to-end autonomous driving architectures.

VADv2 (Default)

Features:

Vector-based scene representation
Planning-oriented perception
Efficient trajectory prediction

Config: configs/worldengine/e2e_vadv2_*.py

Best for: General driving scenarios, fast inference

UniAD

Features:

Unified perception-prediction-planning
Multi-task learning
Strong generalization

Config: configs/worldengine/e2e_uniad_*.py

Best for: Complex scenarios, research

HydraMDP

Features:

Multi-modal trajectory prediction
Distribution-aware planning
Behavior world model integration

Config: configs/worldengine/e2e_hydramdp_*.py

Best for: Safety-critical scenarios, rare cases

Switching Architectures

# Train UniAD instead of VADv2
./scripts/e2e_dist_train.sh configs/worldengine/e2e_uniad_50pct.py 8

# Evaluate HydraMDP
./scripts/e2e_dist_eval.sh \
    configs/worldengine/e2e_hydramdp_50pct.py \
    work_dirs/e2e_hydramdp_50pct/epoch_20.pth \
    8

Advanced Training

Multi-Node Training

For very large models or datasets:

# Node 0 (master)
export MASTER_ADDR=192.168.1.100
export MASTER_PORT=28567
export WORLD_SIZE=16  # Total GPUs
export RANK=0  # Node rank

./scripts/e2e_dist_train.sh configs/worldengine/e2e_vadv2_100pct.py 8

# Node 1 (worker)
export MASTER_ADDR=192.168.1.100
export MASTER_PORT=28567
export WORLD_SIZE=16
export RANK=8

./scripts/e2e_dist_train.sh configs/worldengine/e2e_vadv2_100pct.py 8

Mixed Precision Training

Enable automatic mixed precision (AMP) for faster training:

# In config
fp16 = dict(loss_scale='dynamic')

Gradient Accumulation

For large batch sizes with limited GPU memory:

# In config
data = dict(
    samples_per_gpu=1,
    workers_per_gpu=4,
)

# Set gradient accumulation steps
runner = dict(
    max_epochs=20,
    gradient_accumulation_steps=4,  # Effective batch size = 1 * 8 GPUs * 4 = 32
)

Troubleshooting

Issue 1: CUDA out of memory during training

Solution:

# Reduce batch size
# Edit config: data.samples_per_gpu = 1 (from 2)

# Reduce BEV resolution
# Edit config: bev_h_, bev_w_ = 150, 150 (from 200, 200)

# Use gradient checkpointing
# Edit config: model.img_backbone.with_cp = True

Issue 2: Training loss not decreasing

Possible causes:

Learning rate too high/low
Data loading issues
Incorrect pre-trained weights

Solution:

# Check data loading
python tools/analysis_tools/browse_dataset.py configs/worldengine/e2e_vadv2_50pct.py

# Verify pre-trained weights loaded
grep "load checkpoint" work_dirs/*/logs/train.*

# Try different learning rate
./scripts/e2e_dist_train.sh ... --cfg-options optimizer.lr=1e-4

Issue 3: Evaluation hangs

Solution:

# Check if processes are stuck
ps aux | grep python

# Kill stuck processes
pkill -f "test.py"

# Restart evaluation with fewer GPUs
./scripts/e2e_dist_eval.sh ... 4  # Use 4 instead of 8

Issue 4: "ModuleNotFoundError: No module named mmdet3d"

Solution:

# Ensure you're in the right environment
conda activate algengine

# Verify MMCV installation
python -c "import mmcv; print(mmcv.__version__)"

# Reinstall MMDetection3D if needed
pip uninstall mmdet3d -y
pip install mmdet3d==1.0.0rc6

Issue 5: Checkpoint file corrupted

Solution:

# Use a previous checkpoint
./scripts/e2e_dist_train.sh ... work_dirs/*/epoch_18.pth  # Instead of epoch_20

# Or train from scratch
rm work_dirs/e2e_vadv2_50pct/latest.pth
./scripts/e2e_dist_train.sh ...

Performance Optimization

Training Speed

Increase workers: data.workers_per_gpu = 8 (if CPU/RAM allows)
Use SSD: Store data on fast NVMe SSD
Mixed precision: Enable fp16 = dict(loss_scale='dynamic')
Persistent workers: data.persistent_workers = True

Memory Optimization

Reduce batch size: data.samples_per_gpu = 1
Lower BEV resolution: bev_h_, bev_w_ = 150, 150
Gradient checkpointing: model.img_backbone.with_cp = True
Clear cache: torch.cuda.empty_cache() in code

Distributed Training Tips

Even GPU allocation: Use same GPU type across nodes
InfiniBand: Use high-speed interconnect for multi-node
Shared filesystem: Use NFS/Lustre for data loading
Monitor network: Watch for communication bottlenecks

Next Steps

Run simulations: See SimEngine Usage Guide
Understand evaluation: See Quick Start Guide

For questions, visit GitHub Discussions.

FilesExpand file tree

algengine_usage.md

Latest commit

History

algengine_usage.md

File metadata and controls

AlgEngine Usage Guide

Table of Contents

Quick Reference

Training

Prerequisites

Training from Scratch

Training with 100% Data

Resume Training

Monitor Training

Training Output

Evaluation

Open-Loop Evaluation

Full Test Set Evaluation

Rare Navtest Cases Only

Understanding Evaluation Metrics

Closed-Loop Evaluation

Rare Case Extraction

Basic Extraction

Extracted Splits

Using Custom Thresholds

Verify Extracted Scenarios

Fine-Tuning

Prerequisites: Generating Rollouts with SimEngine

RL-Based Fine-Tuning

Fine-Tuning Configuration

Fine-Tuning Output

Evaluate Fine-Tuned Model

Configuration

Configuration Hierarchy

Key Configuration Parameters

Override Configuration

Model Architectures

VADv2 (Default)

UniAD

HydraMDP

Switching Architectures

Advanced Training

Multi-Node Training

Mixed Precision Training

Gradient Accumulation

Troubleshooting

Issue 1: CUDA out of memory during training

Issue 2: Training loss not decreasing

Issue 3: Evaluation hangs

Issue 4: "ModuleNotFoundError: No module named mmdet3d"

Issue 5: Checkpoint file corrupted

Performance Optimization

Training Speed

Memory Optimization

Distributed Training Tips

Next Steps