Examples

Examples ordered as a learning path — from the end-to-end quickstart to the release gate and real robots. Install per Installation; the Needs column names the uv sync --extra … set, plus any credential or hardware. Time is steady-state wall-clock: the first run of a session additionally pays one-time costs — cold vision-model loads, and for CuRobo examples a one-time CUDA-kernel JIT (~40 s) — so expect minutes before the quoted per-trial rate kicks in.

All examples live under examples/ in the engine repo, each with a README and runnable code.

Start here

Zero to a verified rollout.

Example	What it shows	Needs	Time
libero_quickstart	The end-to-end hero: DINO + SAM3 + VLM perception → OBB grasp → transport, verified against sim ground truth	quickstart + GPU + LLM key	~25–55 s/trial
grocery_packing	A static graph with a real backward edge: perceive → grasp → transport, looped until every item is in the basket, with unprivileged VLM termination	grocery (CUDA) + GPU + LLM key	~min/item

Author & generate graphs

Two producers of the same artifact — gap.builder by hand, the LLM pipeline from one instruction; both pass the same validator and run with the same gap run.

Example	What it shows	Needs	Time
build_a_graph	The full authoring example: 4 subgraphs, a ground-truth checkpoint, recovery actions, optional --execute	uv sync	~1 min
generate_a_graph	Instruction → validated workflow dir; CLI + Python, all providers	uv sync + LLM key	minutes
agent_quickstart	Step-by-step: Claude Code (one skill) generates a graph from a sentence → validate → sim run → video	quickstart + GPU + LLM key + Claude Code	~10 min

Benchmarks & evaluation

The configs that gate releases.

Example	What it shows	Needs	Time
grocery_fulfillment	The flagship acceptance family: graphs LLM-generated per task, nothing hand-written	grocery (CUDA) + LLM key	minutes (smoke)
benchmark	Grid harness configs: 1-cell smoke → position-variance (posvar) grid → the release gate (--gate exits non-zero below the config’s threshold)	grocery (CUDA) + LLM key	minutes → hours

Learned policies

Graphs and policies are complements: graphs steer, collect for, and verify learned policies.

Example	What it shows	Needs	Time
steered_policy	Hybrid graphs: perceive + hover above the target, then hand control to a VLA (vision-language-action) policy	quickstart + policy + LLM key	~min/trial
collect_and_train	Graph as scripted expert → HDF5/LeRobot dataset → train externally → policy back in a graph	quickstart + policy + LLM key	collection: s/episode

Real robots

These move hardware. Read Safety before anything else.

Example	What it shows	Needs	Time
cable_ur	Perception-only UR + ZED connector — motion structurally impossible (read-only RTDE)	real + ZED SDK + UR arm	seconds
real_franka_pick_place	Franka + Robotiq pick-place loop via the robots_realtime bridge	real + Franka/Robotiq/ZED	s/cycle