Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/jackvice/RoboTerrain/llms.txt

Use this file to discover all available pages before exploring further.

RoboTerrain is an open-source framework for developing and evaluating reinforcement learning agents on off-road mobile robots. It bridges ROS 2 Humble, Gazebo Fortress simulation, and PyTorch-backed RL algorithms through a Gymnasium interface — giving you realistic terrain physics, multi-modal sensor observations, and automated navigation metrics out of the box.

Quickstart

Launch the simulation and run your first SAC agent in minutes

RL Environments

Gymnasium environments with LIDAR, IMU, pose, and fused vision observations

Simulation Worlds

Inspection, maze, island, rubicon, and construction Gazebo worlds

Metrics & Analysis

Automated ROS 2 metrics logging and CLI visualization tools

​
What’s Inside

RoboTerrain is structured around three core concerns: high-fidelity simulation, flexible RL training, and rigorous evaluation.

Robot Models

Rover Zero 4WD, Leo Rover, and Clearpath Husky with SDF descriptions

Training Scripts

SAC and PPO agents via Stable Baselines3 with checkpoint callbacks

Active Vision

DreamerV3 agent with joint locomotion and discrete gaze control

Dynamic Obstacles

Spawn pedestrian actors with configurable SDF trajectory files

Docker Setup

Containerized environment with ROS 2 Humble and CUDA support

Reference API

Full environment class API with constructor params and method signatures

​
Research Context

RoboTerrain powers the DUnE benchmark (Dynamic Unstructured Environments) published in Robotics (2025) and the Active Vision for Social Navigation framework. The codebase supports:
  • Point navigation tasks across rough outdoor terrain
  • Social navigation with pedestrian avoidance using attention-based perception
  • Comparative evaluation of SAC, PPO, Nav2 (LiDAR forward/reverse), and Active Vision agents
RoboTerrain requires ROS 2 Humble, Gazebo Fortress, and a CUDA-capable GPU for vision-based training. See the Quickstart for full prerequisites.

Build docs developers (and LLMs) love

Get started for freeTalk to us