Skip to content

directory.md

Latest commit

 

History

History
63 lines (52 loc) · 3.27 KB

directory.md

File metadata and controls

63 lines (52 loc) · 3.27 KB
Raw

Meldog RL — Directory & Path Map

Curated map of the paths agents need: the local repo and the external Isaac Lab reference trees (which agents cannot find by searching this repo). Keep this curated, not exhaustive — update it when structure changes.

Local repo — root /home/frydjak/KNR/Meldog/LRT_meldog_RL

Env + configs

  • source/meldog_rl/envs/meldog_env.py — direct RL env (12-DOF quadruped: 4 legs × T/H/K)
  • source/meldog_rl/envs/__init__.py — gym task registrations (ids below)
  • source/meldog_rl/envs/configs/base_cfg.py — shared base config
  • source/meldog_rl/envs/configs/simulation/{flat,flat_obs,rough,rough_obs}_cfg.py
  • source/meldog_rl/envs/configs/sim2real/* — domain-randomized "Real" variants
  • source/meldog_rl/envs/configs/dataset/* — perception dataset-collection envs
  • source/meldog_rl/envs/configs/simulation/terrain_utils.py
  • source/meldog_rl/agents/rsl_rl_ppo_cfg.py — PPO hyperparameters

Perception

  • source/meldog_rl/models/perception/heightmap_convgru.py — V5 (ConvGRU)
  • source/meldog_rl/models/perception/heightmap_autoreg.py — V6 (autoregressive)
  • source/meldog_rl/models/perception/slam_baseline.py — SLAM baseline
  • source/meldog_rl/models/perception/{projector,voxel_sparse,common}.py

Scripts

  • scripts/locomotion/{train,play,play_keyboard,evaluate}_locomotion.py
  • scripts/perception/{collect_dataset,train_perception,evaluate_perception}.py
  • scripts/perception/{evaluate_slam,debug_cameras}.py
  • scripts/release_model.py — promote a trained run into releases/

Data / outputs / docs

  • logs/locomotion/, logs/perception/ — training runs
  • releases/locomotion/ — promoted checkpoints (v1.0-rough, v1.0-rough-baseline)
  • datasets/ — collected perception datasets
  • thesis/ — LaTeX thesis (chapters/, images/, eiti/)
  • literature/ — reference papers
  • legacy code structure/ — pre-cleanup layout; do not extend
  • AGENTS.md — agent source of truth · agentic/prompts.md — prompts · agentic/manual.md — user manual

Registered task ids

  • Locomotion: Meldog-RL-Locomotion-{Flat,Rough,FlatObs,RoughObs}-{Sim,Real}-v0
  • Dataset: Meldog-RL-Dataset-{Flat,Rough,FlatObs,RoughObs}-v0

External reference — Isaac Lab · root /home/frydjak/IsaacLab

Manager-based velocity locomotion, per-robot config base: source/isaaclab_tasks/isaaclab_tasks/manager_based/locomotion/velocity/config/<robot>/ where <robot> ∈ {anymal_b, anymal_c, anymal_d, go1, go2}. Each has:

  • flat_env_cfg.py — flat-terrain env config
  • rough_env_cfg.py — rough-terrain env config (rewards, obs, terrain, curriculum)
  • agents/rsl_rl_ppo_cfg.py — PPO hyperparameters

Primary references for our tuning:

  • .../config/anymal_c/rough_env_cfg.py — reward terms incl. feet_slide, curriculum
  • .../config/anymal_c/flat_env_cfg.py
  • .../config/anymal_c/agents/rsl_rl_ppo_cfg.py
  • .../config/go2/rough_env_cfg.py — closer scale/dynamics to Meldog
  • .../config/go2/flat_env_cfg.py
  • .../config/go2/agents/rsl_rl_ppo_cfg.py

Direct-workflow ANYmal (matches our direct env style, not manager-based):

  • source/isaaclab_tasks/isaaclab_tasks/direct/anymal_c/anymal_c_env_cfg.py
  • source/isaaclab_tasks/isaaclab_tasks/direct/anymal_c/agents/rsl_rl_ppo_cfg.py