Interaction-Aware Motion Prediction

Learning an interaction-aware motion predictor for autonomous driving: the model forecasts neighboring agents’ future trajectories conditioned on the ego vehicle’s candidate plans, and pairs with a sampling-based planner for collision-aware trajectory selection in SMARTS.

• Video demos: shared folder (OneDrive)

---

Overview

Component	What it does
Interaction-aware predictor	Predicts neighbors’ futures using ego plan + scene context (optional interaction path via --use_interaction).
Sampling-based planner	Generates candidate trajectories, queries the predictor, scores comfort/safety/goal progress, selects a plan.

Decoders (same encoder stack, different recurrent / attention decoder):

Decoder	Module	Notes
LSTM	iamp/predictors/lstm.py	LSTMCell, hidden size 384
GRU	iamp/predictors/gru.py	GRUCell, fewer parameters than LSTM
Transformer	iamp/predictors/transformer.py	Attention-based decoder

Select with --decoder lstm|gru|transformer (no file renaming).

Training scenarios (rotated during training): 1_to_2lane_left_turn_c, 3lane_merge_single_agent, 3lane_overtake.

Metrics: ADE / FDE on predicted trajectories; episode success and collision in logs.

---

Results (Visual)

Unsignalized intersection

Merge

Overtake

---

Installation

1. Prerequisites

• Python 3.9+ (see pyproject.toml). Older SMARTS setups may use 3.7; align versions with your SMARTS install.
• SMARTS comp-1 branch and its system dependencies (SUMO, etc.). This repo includes helper scripts; see smarts.sh.
• GPU recommended for training (PyTorch + CUDA matching your driver).

2. Environment and dependencies

# From the repository root
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

# PyTorch: install the build that matches your CUDA/CPU
# https://pytorch.org/get-started/locally/

pip install -r requirements.txt
pip install -e .

• SMARTS is installed separately (not in requirements.txt). Example:

  pip install "smarts[camera-obs] @ git+https://github.com/huawei-noah/SMARTS.git@comp-1"

  Use the project’s smarts.sh for the full clone + install flow.

• Legacy Docker / pinned stack: source_installations.sh shows an example PyTorch 1.12.0+cu113 install and extra pins used with older environments.

3. Optional: container workflow

run.sh assumes paths like /src/code and sources start.sh, smarts.sh, and source_installations.sh. Adapt paths or run train.py / test.py directly on your machine after SMARTS is installed.

---

Usage

Training

python train.py --use_exploration --use_interaction --decoder lstm

Common flags:

Flag	Meaning
--decoder {lstm,gru,transformer}	Trajectory decoder (default: lstm)
--use_interaction	Interaction-aware predictor path
--use_exploration	Exploration schedule during rollouts
--name NAME	Subfolder under training_log/
--epochs, --batch_size, --learning_rate	Training hyperparameters
--device cuda	Device (default cuda)
--envision_gui, --sumo_gui	GUIs (off by default)

Evaluation

Use the same --decoder as training and point to your checkpoint:

python test.py \
  --model_path models/transformer_1500epochs_predictor_0.6500.pth \
  --use_interaction \
  --decoder transformer \
  --name Test1

Flag	Meaning
--model_path	Path to .pth checkpoint
--decoder	Must match the trained architecture
--episodes	Number of test episodes per scenario

Script driver

Edit run.sh to switch between the train.py and test.py lines as needed.

---

References

• SMARTS