Enhanced Offline Mode

.gitignore

@@ -73,3 +73,6 @@ tests/
 # TacSL sensor
 **/tactile_record/*
 **/gelsight_r15_data/*
+
+# Downloaded assets
+offline_assets/*

CONTRIBUTORS.md

@@ -59,6 +59,7 @@ Guidelines for modifications:
 * Cathy Y. Li
 * Cheng-Rong Lai
 * Chenyu Yang
+* Clayton Littlejohn
 * Connor Smith
 * CY (Chien-Ying) Chen
 * David Yang

scripts/offline_setup/README.md

@@ -0,0 +1,235 @@
+# Isaac Lab Offline Training Setup
+> Complete guide for training Isaac Lab environments offline using locally downloaded assets.
+
+## 🎯 Overview
+#### The offline training system enables you to train Isaac Lab environments without internet connectivity by using locally downloaded assets. This system:
+- ✅ Works with any robot - No hardcoded paths needed
+- ✅ Single flag - Just add --offline to your training command
+- ✅ Automatic fallback - Uses Nucleus if local asset is missing
+- ✅ Maintains structure - Mirrors Nucleus directory organization locally
+
+## 📦 Requirements
+- Isaac Lab installed and working
+- Isaac Sim 5.0 or later
+- 2-20 GB free disk space (depending on assets downloaded)
+- Internet connection for initial asset download
+
+## 🚀 Quick Start
+### 1. Download essential assets (one-time, `all` ~30 GB)
+#### Assets download to the `~/IsaacLab/offline_assets` directory: `cd ~/IsaacLab`
+```
+./isaaclab.sh -p scripts/offline_setup/download_assets.py \
+    --categories all
+```
+#### _Alternative Note: Category fields can be specified separately_
+```
+./isaaclab.sh -p scripts/offline_setup/download_assets.py \
+    --categories Robots --subset Unitree
+```
+### 2. Train completely offline with any robot via the `--offline` flag (also works with `/play`)
+#### Supported for: `rl_games`, `rsl_rl`, `sb3`, `skrl`, and `sim2transfer`
+```
+./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py \
+    --task Isaac-Velocity-Flat-Unitree-Go2-v0 \
+    --num_envs 128 \
+    --offline
+
+./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/play.py \
+    --task Isaac-Velocity-Flat-Unitree-Go2-v0 \
+    --num_envs 128 \
+    --checkpoint logs/rsl_rl/<robot>_flat/<timestamp>/model_<num>.pt \
+    --video \
+    --video_length 1000 \
+    --offline
+```
+#### _Note: For offline training, assets that cannot be found in `offline_assets` will attempted to be fetched from the [Nucleus Server](https://docs.omniverse.nvidia.com/nucleus/latest/index.html)._
+
+## 📁 Asset Layout
+#### Offline assets are organized to mirror Nucleus (`ISAAC_NUCLEUS_DIR` & `ISAACLAB_NUCLEUS_DIR`) under the `offline_assets` directory, meaning that no code changes are required for offline running! We flatten `Isaac/IsaacLab/` to just the category names (`Robots/`, `Controllers/`, etc.) for cleaner local structure. This happens in `asset_resolver.py`, where the resolver maintains a 1:1 mapping between Nucleus and local storage.
+
+```
+IsaacLab/
+├── source/isaaclab/isaaclab/utils/
+│   └── asset_resolver.py           # Core resolver
+├── scripts/setup/
+│   └── download_assets.py          # Asset downloader
+└── offline_assets/
+    ├── ActuatorNets/
+    ├── Controllers/
+    ├── Environments/               # Ground planes
+    │   └── Grid/
+    │       └── default_environment.usd
+    ├── Materials/                  # Textures and HDRs
+    │   └── Textures/
+    │       └── Skies/
+    ├── Mimic/
+    ├── Policies/
+    ├── Props/                      # Markers and objects
+    │   └── UIElements/
+    │       └── arrow_x.usd
+    └── Robots/                     # Robot USD files
+        ├── BostonDynamics/
+        │   └── spot/
+        │       └── spot.usd
+        └── Unitree/
+            ├── Go2/
+            │   └── go2.usd
+            └── H1/
+                └── h1.usd
+```
+
+
+
+
+
+
+
+
+
+### Proposal
+
+Add offline training support to Isaac Lab that enables training environments without internet connectivity. The feature should automatically redirect asset paths from Nucleus/S3 servers to local storage while maintaining the same directory structure, requiring zero configuration changes to existing environments.
+
+Core capabilities:
+- One-time asset download script that mirrors Nucleus directory structure locally
+- Automatic path resolution from Nucleus URLs to local filesystem
+- Single `--offline` flag for all training scripts
+- Graceful fallback to Nucleus if local asset is missing
+- Works with any robot and environment without hardcoded paths
+
+### Motivation
+
+**Current Problem:**
+Training Isaac Lab environments requires constant internet connectivity to load assets from Nucleus/S3 servers. This creates several critical issues:
+
+1. **Airgapped Environments**: Impossible to train in secure/classified facilities that prohibit internet access
+2. **Network Reliability**: Training fails or becomes extremely slow on unstable networks
+3. **Reproducibility**: Repeated downloads slow iteration and depend on external server availability
+4. **Development Workflow**: Researchers waste time waiting for assets during rapid prototyping
+5. **DGX Spark is Rooted**: I should be able to take my Spark anywhere regardless of internet connectivity and live out Isaac Lab and Isaac Sim to it's fullest
+
+**User Story:**
+"I'm always frustrated when I need to train in an airgapped lab environment or on an unstable network connection. I have to create custom configs with hardcoded local paths for each robot, which results in unmaintainable code duplication and can break when I switch robots or update Isaac Lab. I would like to take Isaac Lab on the go and quickly demo anything in an independent localized ecosystem without nit-picking configurations and dealing with asset management."
+
+**Core User Needs:**
+- Train without internet connectivity
+- Avoid hardcoded paths in environment configs
+- Work seamlessly with any robot
+- Maintain same configs for both online and offline modes
+
+### Alternatives
+
+**Alternative 1: Hardcoded Local Paths** (Current workaround)
+```python
+# Separate config file per robot
+UNITREE_GO2_CFG = ArticulationCfg(
+    spawn=sim_utils.UsdFileCfg(
+        usd_path=f"{ISAACLAB_NUCLEUS_DIR}/Robots/Unitree/Go2/go2.usd"
+```
+**Problems:**
+- ❌ Requires separate config for each robot
+- ❌ ~30 lines of boilerplate per robot
+- ❌ Breaks when Isaac Lab updates
+- ❌ Not maintainable
+
+**Alternative 2: Environment Variables**
+```bash
+export ISAAC_ASSETS_DIR="/local/path"
+```
+**Problems:**
+- ❌ Doesn't work with S3 URLs
+- ❌ Requires Isaac Sim core changes
+- ❌ Can't have fallback to Nucleus
+- ❌ Not transparent to users
+
+**Alternative 3: Manual Asset Copying**
+Copy assets manually and update configs.
+**Problems:**
+- ❌ Error-prone manual process
+- ❌ No directory structure guidelines
+- ❌ Still requires config changes
+
+**Proposed Solution Benefits:**
+- ✅ Zero code or config changes
+- ✅ Works with all robots and environments automatically
+- ✅ Single flag: `--offline` for all training environments (`rl_games`, `rsl_rl`, `sb3`, `skrl`, and `sim2transfer`)
+- ✅ Automatic fallback
+- ✅ Easy `offline_asset` plug and play following Nucleus structure
+- ✅ 90% code reduction
+
+### Build Info
+
+- Isaac Lab Version: main branch (as of January 2026)
+- Isaac Sim Version: 5.1.0
+
+### Additional Context
+
+**Use Cases:**
+1. **Defense/Aerospace**: Training in classified airgapped facilities
+2. **Remote Locations**: Field robotics research with limited connectivity
+3. **Development**: Rapid iteration without network delays and firewall interruptions
+4. **CI/CD**: Reproducible builds without external dependencies
+5. **Workshops/Tutorials**: Teaching without relying on conference WiFi (i.e. localized everything on DGX Spark)
+
+**Technical Approach:**
+The implementation uses a dual-layer strategy:
+- **Monkey patching**: Intercepts asset loads at spawn config instantiation (90% coverage)
+- **Config patching**: Explicitly modifies pre-loaded configs (10% coverage)
+
+This ensures ~100% asset coverage without modifying environment configs.
+
+**Expected Directory Structure:**
+```
+IsaacLab/
+├── source/isaaclab/isaaclab/utils/
+│   └── asset_resolver.py           # Core resolver
+├── scripts/setup/
+│   └── download_assets.py          # Asset downloader
+└── offline_assets/
+    ├── ActuatorNets/...
+    ├── Controllers/...
+    ├── Environments/               # Ground planes
+    │   └── Grid/
+    │       └── default_environment.usd
+    ├── Materials/                  # Textures and HDRs
+    │   └── Textures/
+    │       └── Skies/
+    ├── Mimic/...
+    ├── Plocies/...
+    ├── Props/                      # Markers and objects
+    │   └── UIElements/
+    │       └── arrow_x.usd
+    └── Robots/                     # Robot USD files
+        ├── BostonDynamics/
+        │   └── spot/
+        │       └── spot.usd
+        └── Unitree/
+            ├── Go2/
+            │   └── go2.usd
+            └── H1/
+                └── h1.usd
+```
+
+Dynamically pulls and mirrors Nucleus structure for seamless path resolution.
+
+### Checklist
+
+- [x] I have checked that there is no similar issue in the repo (**required**)
+
+### Acceptance Criteria
+
+- [x] Asset download script that mirrors Nucleus directory structure to local storage (offline_assets)
+- [x] Automatic path resolver that redirects Nucleus URLs to local filesystem
+- [x] Optional `--offline` flag added to all training scripts (RSL-RL, SB3, SKRL, RL Games)
+- [x] Monkey patching of Isaac Lab spawn configs (UsdFileCfg, GroundPlaneCfg, PreviewSurfaceCfg)
+- [x] Config patching for pre-loaded environment configs
+- [x] Graceful fallback to Nucleus for missing assets
+- [x] Support for versioned Nucleus URLs (e.g., `/Assets/Isaac/5.1/...`)
+- [x] Documentation including setup guide, usage examples, and troubleshooting
+- [x] Works with any robot without hardcoded paths
+- [x] Zero breaking changes - existing code continues to work
+- [x] Manual testing completed across multiple robots (Go2, H1, ANYmal)
+- [x] Verification in complete offline mode (no internet connectivity)
+
+**Definition of Done:**
+A user can download all, or select assets by running `./isaaclab.sh -p scripts/offline_setup/download_assets.py --categories all`, then train any robot completely offline by simply adding `--offline` to their training command, with no code or config changes required.