Demos

This section captures the canonical shell utilities used to reproduce MarineGym benchmark experiments. The batch runner iterates across vehicle variants, task configurations, and disturbance regimes to generate a complete suite of training and evaluation runs.

For a concise command primer, see Quick Start. The recipes below focus on automating large experiment batches and documenting coverage.

Batch Runner Overview

The script below launches 45 workflows (5 robots × 3 tasks × 3 modes) with shared hyperparameters:

• TOTAL_FRAMES: 50,000,000

• SEED: 47

• task.env.num_envs: 2048

• algo: Proximal Policy Optimization (PPO)

Each job sleeps for five seconds before the next iteration to avoid scheduler contention.

#!/bin/bash

TOTAL_FRAMES=50000000
SEED=47

ROBOTS=("HAUV" "iAUV" "LAUV" "BlueROV" "BlueROVHeavy")
TASKS=("TrackRand" "LandingRand" "HoverRand")
MODES=("st" "dt" "dr")

run_training() {
    local robot=$1
    local task=$2
    local mode=$3

    echo "Running: ${robot} / ${task} / ${mode}"

    case $mode in
        st)
            python train.py task=$task \
                algo=ppo \
                headless=true \
                enable_livestream=false \
                mode=train \
                task.drone_model.name=$robot \
                task.env.num_envs=2048 \
                total_frames=$TOTAL_FRAMES \
                seed=$SEED
            ;;
        dt)
            python train.py task=$task \
                algo=ppo \
                headless=true \
                enable_livestream=false \
                mode=train \
                task.drone_model.name=$robot \
                task.disturbances.train.flow.enable_flow=true \
                task.disturbances.train.payload.enable_payload=false \
                task.env.num_envs=2048 \
                total_frames=$TOTAL_FRAMES \
                seed=$SEED
            ;;
        dr)
            python train.py task=$task \
                algo=ppo \
                headless=true \
                enable_livestream=false \
                mode=train \
                task.drone_model.name=$robot \
                task.disturbances.train.flow.enable_flow=true \
                task.disturbances.train.payload.enable_payload=true \
                task.env.num_envs=2048 \
                total_frames=$TOTAL_FRAMES \
                seed=$SEED
            ;;
    esac

    echo "Completed: ${robot} / ${task} / ${mode}"
    sleep 5
}

for robot in "${ROBOTS[@]}"; do
    for task in "${TASKS[@]}"; do
        for mode in "${MODES[@]}"; do
            run_training "$robot" "$task" "$mode"
        done
    done
done

Mode Reference

Mode	Purpose	Additional arguments
st	Baseline training without disturbance	(Uses defaults)
dt	Disturbance-aware training	task.disturbances.train.flow.enable_flow=true task.disturbances.train.payload.enable_payload=false
dr	Disturbance-aware training with randomization	task.disturbances.train.flow.enable_flow=true task.disturbances.train.payload.enable_payload=true

Robots × Tasks Matrix

The grid below enumerates all combinations executed by the batch script. Each cell yields three individual runs (st, dt, dr).

Scenario coverage

Robot	TrackRand	LandingRand	HoverRand
HAUV	st, dt, dr	st, dt, dr	st, dt, dr
iAUV	st, dt, dr	st, dt, dr	st, dt, dr
LAUV	st, dt, dr	st, dt, dr	st, dt, dr
BlueROV	st, dt, dr	st, dt, dr	st, dt, dr
BlueROVHeavy | st, dt, dr		st, dt, dr	st, dt, dr

Single-Run Templates

Use the following snippets when launching an individual job outside the batch loop. Replace <ROBOT> and <TASK> with the desired values.

Standard training

python train.py task=<TASK> \
    algo=ppo \
    headless=true \
    enable_livestream=false \
    mode=train \
    task.drone_model.name=<ROBOT> \
    task.env.num_envs=2048 \
    total_frames=50000000 \
    seed=47

Disturbance training

python train.py task=<TASK> \
    algo=ppo \
    headless=true \
    enable_livestream=false \
    mode=train \
    task.drone_model.name=<ROBOT> \
    task.disturbances.train.flow.enable_flow=true \
    task.disturbances.train.payload.enable_payload=false \
    task.env.num_envs=2048 \
    total_frames=50000000 \
    seed=47

Disturbance and randomization training

python train.py task=<TASK> \
    algo=ppo \
    headless=true \
    enable_livestream=false \
    mode=train \
    task.disturbances.train.flow.enable_flow=true \
    task.disturbances.train.payload.enable_payload=true \
    task.drone_model.name=<ROBOT> \
    task.env.num_envs=2048 \
    total_frames=50000000 \
    seed=47