Meldog — RL Perception and Locomotion

End-to-end framework for developing locomotion and perception policies for a quadruped robot Meldog using reinforcement learning and supervised learning.

Demo

Locomotion Policy

Meldog_Locomotion.mp4

Terrain Perception

Meldog_Perception.1.mp4

Overview

This project provides a complete pipeline for training locomotion and perception:

1. Train locomotion policy — RL-based controller (PPO) that learns to walk on varied terrain
2. Collect perception dataset — use trained locomotion to gather depth camera observations with ground truth height maps
3. Train perception model — supervised learning to reconstruct dense terrain from sparse observations

Perception Models

Currently there are three functioning versions of the terrain perception network:

Input: Sparse height map (40×40) from 4 depth cameras + gravity vector from IMU
Output: Reconstructed height map (40×40)

V1 — Base Model

Standard U-Net encoder-decoder with skip connections. Gravity vector is embedded via MLP and injected at the bottleneck.

V2 — Deep + Attention

Deeper encoder (additional level down to 5×5) with self-attention mechanism. Larger receptive field allows the network to capture global terrain structure.

V3 — Temporal

Adds ConvGRU layers that maintain hidden state across frames. The network builds a "belief state" about terrain — remembering previously observed regions that are now occluded.

Project Structure

Built on Isaac Lab's direct workflow template. Here's where key components live:

Training Pipelines

• scripts/locomotion/ — RL training, evaluation, and testing for locomotion policies
• scripts/perception/ — Supervised learning pipeline for terrain perception (training, evaluation, dataset collection)

Robot Description

• source/meldog_rl/envs/meldog_env.py — Main environment implementation defining robot physics, observations, rewards
• source/meldog_rl/envs/configs/ — Environment configurations organized by use case:
  • simulation/ — Training configs (flat/rough terrain)
  • sim2real/ — Real-world deployment configs with domain randomization
  • dataset/ — Dataset collection configs with camera observations enabled

Neural Network Models

• source/meldog_rl/models/perception/ — Perception architectures (V1 base, V2 attention, V3 temporal)
• source/meldog_rl/agents/ — Locomotion policy configuration (PPO hyperparameters, network architecture)

Data

• logs/locomotion/ — RL training checkpoints and TensorBoard logs
• logs/perception/ — Perception model checkpoints and training metrics
• datasets/ — Collected depth camera observations with ground truth height maps

Supporting Code

• source/meldog_rl/datasets/ — PyTorch dataset classes for perception training
• source/meldog_rl/utils/ — Shared utilities and naming conventions

Installation

This project is built on the Isaac Lab template.

TBD

Usage

TBD

Acknowledgments

Built using Isaac Lab simulation framework and project template.