RoboLab Documentation

How RoboLab Works

RoboLab dynamically combines tasks with user-specified robot, observations, actions, and simulation parameters at environment registration time.

Terminology

Term	Meaning
scene	A USD/USDA file describing the static contents of a workspace — objects, fixtures, table, and their spatial layout. Reusable across tasks. See Scenes.
task	A `Task` dataclass binding a scene to a language instruction, termination criteria, and (optional) subtasks. See Tasks.
environment	A task combined with robot, camera, lighting, background, and simulation configs, registered as a Gymnasium env. `--num-envs N` spawns `N` parallel instances in a grid, each indexed by `env_id`. See Environment Registration.
episode	One trajectory from one instance of an environment from reset to termination.
run	One sequential pass over all environments (one reset → step loop → termination → `end_episode` cycle). If running with `--num-envs N`, then each run produces `N` episodes.

The core concepts are:

Objects, Scenes, Tasks

Objects — USD object assets with physics properties for manipulation
Scenes — USD-based environments containing objects, fixtures, and spatial layout
Tasks — Language instructions, termination criteria, and scene bindings
Task Libraries — Managing task collections, generating metadata, and viewing statistics

Task Conditionals

Subtask Checking — Granular progress tracking within tasks
Conditionals — Predicate logic for defining success/failure conditions
Event Tracking — Monitoring task-relevant events during execution

Variations

Robots — Robot articulation configs, actuators, and action spaces
Cameras — Scene cameras and robot-attached cameras
Lighting — Scene lighting (sphere, directional, and custom lights)
Backgrounds — HDR/EXR dome light backgrounds

Environments

Environment Registration — How tasks are combined with robot/observation/action configs into runnable Gymnasium environments
Environment Generation — Contact sensor creation, subtask trackers, and runtime environment internals
Running Environments — Creating environments, evaluation scripts, CLI reference, and robustness testing
num_envs VRAM size guide — Per-task num_envs ceiling on L40, measured against pi05

Policy

Inference Clients — Built-in policy clients and server setup instructions

Output

Data Storage and Output — Output directory structure, HDF5 layout, and episode result fields
Analysis and Results Parsing — Scripts for summarizing, comparing, and auditing experiment results

Debug

Debugging — Verbose/debug flags, world state inspection, and diagnostic scripts
Known Issues — Documented bugs and workarounds

Developing and Working with RoboLab

If you're building a new benchmark and a new experiment workflow, follow the steps below in order. Otherwise, pick whichever applies to your use case.

Creating new assets, tasks, and benchmarks

Creating New Objects — Author USD object assets with rigid body, collision, and friction properties. Includes pipeline for catalog generation, screenshots, and physics tuning.
Creating New Scenes — Compose objects into USD scenes using IsaacSim. Includes settling, metadata generation, and screenshot utilities.
Creating New Tasks — Define task dataclasses with language instructions, termination criteria, and scene bindings. Tasks can live in your own repository.
Managing Task Libraries — Organize tasks into collections, generate metadata (JSON, CSV, README), and compute statistics.

Configuring robots, cameras, lighting, and backgrounds

Robots — Define or customize robot articulation, actuators, and action spaces. Use built-in configs (DROID, Franka) or bring your own from IsaacLab.
Cameras — Set up scene cameras and robot-attached cameras (e.g., wrist cameras).
Lighting — Configure scene lighting for evaluation or robustness testing.
Backgrounds — Set HDR/EXR dome light backgrounds for realistic scene rendering.

Evaluating a new policy against the benchmark

Setting Up Environment Registration — Register tasks with your robot/observation/action/simulation settings. For DROID with joint-position actions, the built-in registration can be used directly.
Evaluating a New Policy — Implement an inference client for your model and run multi-task evaluation. Everything can live in your own separate repository.

Analysis

Statistical Significance of Results — Discussion on how to run evaluations such that your results are statistically significant.

Browsing the benchmark and eval results

Dashboard — Self-contained web viewer for scenes, tasks, and eval outputs. Runs locally with robolab-dashboard --output-dir output/; binds 0.0.0.0 so anyone on your LAN can hit your IP.

AI Workflows

Scene Generation — Generate USD scenes from natural language using the /robolab-scenegen Claude Code skill. See skills/robolab-scenegen/.
Task Generation — Generate task files from natural language using the /robolab-taskgen Claude Code skill. See skills/robolab-taskgen/.