DETB

DETB is a simulation-first design evaluation toolkit for quadruped systems engineering. It is designed to turn Isaac Lab experiments into reviewable engineering evidence, not deployment claims or sim-to-real guarantees.

What DETB Is

• A reproducible experiment runner for quadruped design studies.
• A contract-first pipeline for manifests, metrics, and artifact bundles.
• A framework for terrain, sensor, fault, and tuning studies.
• A public-facing GitHub repository intended for technical review.

What DETB Is Not

• A field deployment stack.
• A certification tool.
• A claim of validated sim-to-real transfer.
• A generic reinforcement learning benchmark zoo.

V1 Support Contract

The concrete v1 operator path is:

• train
• evaluate
• visualize
• bundle-artifacts
• generate-requirements

The v1 contract is supported on the mock backend for fast local validation, and the real Isaac Lab backend is verified for the core simulator-backed path that feeds those same artifacts.

The following commands are experimental and are not part of the v1 support contract:

• train-gui
• sweep
• sensor-eval
• terrain-eval
• failure-eval
• tune

Current Status

As of 2026-04-14, the repository supports the v1 operator contract above with a complete mock backend for fast local validation and a verified Isaac Lab backend for the core simulator-backed flows. Simulator-native task and robot ownership live under source/detb_lab, while detb/ remains the orchestration and evidence layer. Windows GUI launches through DETB use --/app/vulkan=false on this machine so the GUI path stays stable and reviewable in the launch spec.

Pinned Baseline

• Isaac Sim: 5.1.0
• Isaac Lab checkout: ..\IsaacLab-5.1
• Isaac Lab version file: 2.3.0
• Conda environment: isaaclab51
• Default device: cuda:0
• Default GPU index: 0

On 2026-04-14, GPU 0 on this machine resolved to NVIDIA GeForce RTX 5080 Laptop GPU with driver 595.79. DETB records the detected GPU and driver automatically in the run manifest.

Session Setup

For low-friction sessions, see docs/permissions.md. The practical target is one persistent approval for python -m detb.cli plus one for python -m pytest.

Quick Start

cd "C:\Dev Projects\NVIDIA Omniverse\IsaacLab Testbed"
python -m pip install -e .[dev]
python -m pip install -e source/detb_lab
detb train
detb evaluate
detb visualize --set visualization.execute=true
detb bundle-artifacts --source-dir outputs/evaluate/<run_id>
detb generate-requirements --source-dir outputs/evaluate/<run_id>

Experimental Workflows

The following commands are useful for development, but they are not part of the v1 support contract:

• train-gui
• sweep
• sensor-eval
• terrain-eval
• failure-eval
• tune

Use dry-run launch specs first so the exact Isaac Lab command is reviewable before anything opens:

detb visualize
detb visualize --set visualization.execute=true

For the baseline task=flat_walk path, visualization.use_pretrained_checkpoint=true now resolves to the upstream Isaac Lab ANYmal-C pretrained checkpoint. DETB-only task variants without a published upstream equivalent fail fast instead of silently replaying the latest local smoke checkpoint.

The GUI training path remains available for manual inspection, but it is experimental:

detb train-gui
detb train-gui --set visualization.train_execute=true

Tune and study workflows are also experimental:

detb tune --set study=sweep
detb tune --set study=sweep --set objective.terrain_weight=0.5 --set objective.target_tgs=0.8
detb sweep
detb sensor-eval
detb terrain-eval
detb failure-eval

Select a DETB-owned task variant directly from Hydra when you want a different Isaac registry family:

detb train --set task=flat_walk_stability
detb evaluate --set task=flat_walk_stability
detb visualize --set task=flat_walk_stability --set visualization.execute=true

Select the simple-actuator robot/profile path by pairing the matching task and robot groups:

detb train --set task=flat_walk_simple_actuator --set robot=anymal_c_simple_actuator
detb evaluate --set task=flat_walk_simple_actuator --set robot=anymal_c_simple_actuator

Supported Commands

Command	Purpose
detb train	Produce a manifest, training summary, reward curve artifacts, and either a synthetic checkpoint (mock) or copied real checkpoint/config snapshots (isaaclab).
detb evaluate	Produce episode metrics, aggregate metrics, overview plot, summary markdown, and Isaac run metadata when execution.backend=isaaclab.
detb visualize	Launch the pinned Isaac Lab GUI playback path or emit a launch spec when visualization.execute=false.
detb bundle-artifacts --source-dir ...	Rebuild the summary and bundle artifacts from an existing run directory.
detb generate-requirements --source-dir ...	Generate a candidate requirement ledger from stored artifacts.

Experimental Commands

Command	Purpose
detb train-gui	Launch the pinned Isaac Lab GUI training path or emit a launch spec when visualization.train_execute=false.
detb sweep	Run staged design-point screening using the configured study points.
detb sensor-eval	Compare configured sensor profiles under matched conditions.
detb terrain-eval	Run the configured terrain battery and compute TGS.
detb failure-eval	Sweep fault severity and detect the first critical threshold.
detb tune	Rank design points using objective weights and target thresholds.

Artifact Contract

Each major run writes to outputs/<command>/<run_id>/ and includes:

• resolved_config.yaml
• run_manifest.json
• artifact_registry.json
• machine-readable CSV or JSON outputs
• reviewable markdown summary
• lightweight SVG plots where applicable

GUI commands also emit launch-spec JSON files so the exact Isaac Lab invocation is reviewable before execution. Executed visualize runs now persist isaac_play_result.json, playback_telemetry.csv, playback logs, and any recorded videos/play/*.mp4 files in the same run directory.

Repository Layout

configs/      Hydra configs and pinned defaults
detb/         Core package, runtime, backends, and contracts
source/detb_lab/  External Isaac Lab extension package for custom tasks and robot configs
docs/         Bootstrap, architecture, research, and workflow notes
scripts/      Convenience launchers
tests/        Unit and integration tests
outputs/      Generated artifacts (gitignored)

Documentation

• docs/bootstrap.md
• docs/architecture.md
• docs/data-contracts.md
• docs/cli.md
• docs/modules.md
• docs/reproducibility.md
• docs/terrain-research.md
• docs/terrain-design.md

Roadmap

1. Keep the v1 operator contract stable across train, evaluate, visualize, bundle-artifacts, and generate-requirements.
2. Keep the mock-backed pipeline stable as the contract baseline.
3. Expand DETB-owned task families and robot configs under source/detb_lab.
4. Keep experimental study commands clearly separated from the v1 support contract.
5. Continue using GUI launch specs and Isaac Lab playback/training to inspect progress visually when needed.
6. Expand the verified real Isaac Lab backend only after the v1 path remains stable.