PGTT: Phase-Guided Terrain Traversal for Perceptive Legged Locomotion

A perceptive reinforcement learning locomotion framework developed for the Unitree GO2 and ANYmal in simulation (MuJoCo MJX). Deployed on real hardware using unitree sdk2py for robot control and Point-Lio and Elevation Mapping for the perception pipeline, using the Unitree L1 LiDAR.

Project Structure

.
├── robots/                  # Shared robot-agnostic code
│   ├── __init__.py          # Robot registry: robots.get_robot_config("go2")
│   ├── base.py              # RobotEnv base class (parameterized by consts)
│   ├── joystick_base.py     # Shared Joystick_Base with all reward functions
│   ├── joystick_pgtt.py     # PGTT joystick variant
│   ├── joystick.py          # Baseline joystick variant
│   ├── joystick_wild.py     # Wild joystick variant (oscillator pose)
│   ├── gait.py              # Gait controller (spline-based trajectories)
│   ├── heightmap.py         # Heightmap sensor (JAX, parameterized grid spacing)
│   ├── randomize.py         # Domain randomization (parameterized offsets)
│   ├── randomize_simple.py  # Simple domain randomization
│   └── utility.py           # Shared utilities
├── go2/
│   ├── robot_config.py      # Go2-specific constants, configs, and defaults
│   └── xmls/                # Go2 MuJoCo XML scene files
├── anymal/
│   ├── robot_config.py      # ANYmal-specific constants, configs, and defaults
│   └── xmls/                # ANYmal MuJoCo XML scene files
├── training/
│   └── train.py             # Unified training script (--robot flag)
├── deploy/
│   ├── deploy_heightmap.py  # MuJoCo simulation deployment (--robot flag)
│   ├── deploy_real.py       # Real hardware deployment (Unitree SDK)
│   ├── cpu_heightmap/       # CPU-based heightmap for deployment
│   └── policy_net.py        # Policy network loader (PyTorch)
├── terrain/
│   └── generator.py         # Terrain generation via WFC (--robot flag)
└── policies/                # Saved policy checkpoints

All shared logic lives in robots/. Robot-specific parameters (XML paths, PD gains, heightmap spacing, reward scales, training hyperparameters) are defined in go2/robot_config.py and anymal/robot_config.py. Every script accepts a --robot flag to select which robot to use.

Terrain Generation

Terrains are produced using Wave Function Collapse in MuJoCo.

To generate terrains:

# Generate level .npy files for training (Go2 default)
python terrain/generator.py levels

# Generate level .npy files for ANYmal
python terrain/generator.py --robot anymal levels

# Fill template XML with placeholder boxes
python terrain/generator.py fill --num_objects 100

# Create a random test terrain for visualization
python terrain/generator.py test --step_height 0.08 --width 0.4 --num_steps 4

See terrain/README.md for the full argument reference for each subcommand.

Training Pipeline

Real World Experiment

The resulting policy deployed in the real world. We use Point-Lio for odometry and Gridmap to extract the desired heightmap.

pgtt_real.mp4

Installation

Create a conda environment (recommended):

conda create -n pgtt python=3.12 -y
conda activate pgtt

Install JAX for GPU for your CUDA version. To find your CUDA version:

nvidia-smi

pip install "jax[cuda12]==0.8.0"

if CUDA version is 13.

Install the required dependencies:p

pip install -r requirements.txt

Training

Quick start

# Train Go2 with PGTT method on stairs
python training/train.py --robot go2 --method pgtt --task_name stairs

# Train ANYmal with PGTT method on stairs
python training/train.py --robot anymal --method pgtt --task_name stairs

Available methods

Method	Description
pgtt	Phase-Guided Terrain Traversal (default)
baseline	Baseline without phase/gait frequency in obs
wild	Uses oscillator pose from gait trajectory generator

Training arguments

Argument	Default	Description
--robot	go2	Robot: go2 or anymal
--method	pgtt	Training method: pgtt, baseline, or wild
--task_name	stairs	Task: stairs or flat_terrain
--terrain_file	terrains/level04.npy	Path to terrain heightmap file
--num_envs	from robot config	Number of parallel environments
--batch_size	from robot config	PPO batch size
--num_minibatches	from robot config	Number of PPO minibatches
--num_timesteps	200_000_000	Total training timesteps
--learning_rate	3e-4	Learning rate
--discount	0.97	Discount factor
--num_evals	31	Number of evaluations during training
--index	0	Identifier for checkpoint saving
--checkpoint_folder	None	Resume training from checkpoint
--eval_flag	False	Enable evaluation mode (restricted commands)

Default training hyperparameters (num_envs, batch_size, num_minibatches, gait_freq) are loaded from each robot's config and can be overridden via CLI.

Examples

# Train Go2 baseline on flat terrain
python training/train.py --robot go2 --method baseline --task_name flat_terrain

# Train ANYmal PGTT with custom hyperparameters
python training/train.py --robot anymal --method pgtt --num_envs 4096 --batch_size 1024

# Resume training from checkpoint
python training/train.py --robot go2 --method pgtt --checkpoint_folder checks_stairs/checkpoint_0

Full training pipeline (training.sh)

The main training script runs all methods across multiple terrain difficulty levels with curriculum learning. It takes the robot name as the first argument:

# Train Go2 (default)
bash training/training.sh go2

# Train ANYmal
bash training/training.sh anymal

This runs 5 independent runs, each training pgtt, baseline, and wild methods across levels level03 → level07 → level10 → level13, using curriculum learning (each level resumes from the previous checkpoint).

Evaluation

Evaluate trained checkpoints on discrete terrains and/or stair heights:

# Run both discrete and stair evaluations for Go2
python training/evaluate_multiple.py --robot go2 --eval_type both

# Only discrete terrain evaluation for ANYmal
python training/evaluate_multiple.py --robot anymal --eval_type discrete

# Only stair height evaluation with custom height range
python training/evaluate_multiple.py --robot go2 --eval_type stairs --height_min 1 --height_max 10

# Evaluate a specific training run
python training/evaluate_multiple.py --robot go2 --eval_type both --run 2

Argument	Default	Description
--robot	go2	Robot: go2 or anymal
--eval_type	both	Evaluation type: stairs, discrete, or both
--run	0	Which training run to evaluate
--height_min	1	Min stair height in cm (for stairs eval)
--height_max	10	Max stair height in cm (for stairs eval)

Results are saved as .npy files in plots/.

Deployment

Simulation deployment (MuJoCo viewer)

# Deploy Go2 policy in simulation
python deploy/deploy_heightmap.py --robot go2

# Deploy ANYmal policy in simulation
python deploy/deploy_heightmap.py --robot anymal

# Enable or disable stairs terrain
python deploy/deploy_heightmap.py --stairs
python deploy/deploy_heightmap.py --no-stairs

# Override command velocity
python deploy/deploy_heightmap.py --vx 1.0 --vy 0.0 --yaw 0.0

Default deployment parameters (command velocity, gait frequency, policy path, perturbation settings) are loaded from DEPLOY_DEFAULTS in each robot's config. CLI flags override the config values; any flag that is omitted falls back to the config default.

Argument	Default	Description
--robot	go2	Robot: go2 or anymal
--method	pgtt	Method: pgtt, baseline, or wild
--level	level03	Terrain level used to select the policy
--run	0	Policy run index
--stairs / --no-stairs	from config	Enable or disable stair terrain
--vx	from config	Forward speed command (m/s)
--vy	from config	Lateral speed command (m/s)
--yaw	from config	Yaw rate command (rad/s)

Real hardware deployment

Two terminals are required:

Terminal 1 — Perception pipeline (Docker)

Make sure /utlidar/cloud and /utlidar/imu topics are providing data, then start the elevation mapping container (see Perception pipeline):

docker run -it --rm \
  -e DISPLAY=$DISPLAY \
  --net=host \
  -e LIBGL_ALWAYS_SOFTWARE=1 \
  -e MESA_LOADER_DRIVER_OVERRIDE=llvmpipe \
  -v /tmp/.X11-unix:/tmp/.X11-unix:rw \
  ros2-humble-dev

Inside the container(only when you actually want to start the heightmap and the robot is in stand mode):

cd ros_scripts
chmod +x run_elevation.sh
./run_elevation.sh

An RViz window should appear showing the gridmap and scan points. Leave this running.

Terminal 2 — Policy (conda env pgtt)

conda activate pgtt

# Deploy with controller (joystick) commands (default)
python -m deploy.deploy_real --robot go2 --method pgtt --level level03 --run 0

# Deploy with fixed velocity commands
python -m deploy.deploy_real --robot go2 --method pgtt --level level03 --run 0 \
    --command_type fixed --vx 0.2 --vy 0.0 --yaw 0.0

Argument	Default	Description
--robot	go2	Robot: go2 or anymal
--method	pgtt	Method: pgtt, baseline, or wild
--level	level03	Terrain level used to select the policy checkpoint
--run	0	Policy run index
--command_type	controller	Command source: controller (joystick) or fixed
--vx	0.2	Forward speed when using fixed command (m/s)
--vy	0.0	Lateral speed when using fixed command (m/s)
--yaw	0.0	Yaw rate when using fixed command (rad/s)

Startup sequence:

1. Press Enter in Terminal 2 to confirm and begin.
2. The robot automatically stands up over ~7 seconds.
3. Once fully standing, the terminal prints "Press B to start policy execution.".Switch to Terminal 1 (Docker) and run ./run_elevation.sh. Wait for RViz to show a stable, drift-free heightmap with correct ground estimation.
4. Press B on the controller to start the policy.
5. Press A at any time to immediately stop the policy and sit the robot down safely.

Adding a New Robot

1. Create a directory <robot_name>/ with xmls/ containing your MuJoCo scene files.
2. Create <robot_name>/robot_config.py following the structure of go2/robot_config.py:
   • Define XML paths, sensor names, heightmap spacing, PD gains
   • Define default_config(), baseline_config(), wild_config() functions
   • Define TRAINING_DEFAULTS and DEPLOY_DEFAULTS dicts
3. Register the robot in robots/__init__.py:

   _REGISTRY = {
       "go2": "go2.robot_config",
       "anymal": "anymal.robot_config",
       "your_robot": "your_robot.robot_config",
   }

4. All training, deployment, and terrain scripts will work with --robot your_robot.

Perception pipeline using docker

Build the image once:

docker build -t ros2-humble-dev .

Running instructions are part of the Real hardware deployment two-terminal setup above.

Citation

If you find our work useful in your research, please cite it as follows:

@inproceedings{ntagkas2025pgtt,
  title={PGTT: Phase-Guided Terrain Traversal for Perceptive Legged Locomotion},
  author={Ntagkas, Alexandros and Kiourt, Chairi and Chatzilygeroudis, Konstantinos},
  booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2026}
}